GT English

GRAPH THEORY
Keijo Ruohonen
(Translation by Janne Tamminen, Kung-Chung Lee and Robert Pich)
2008
Contents
1 I DEFINITIONS AND FUNDAMENTAL CONCEPTS
1 1.1 Denitions
6 1.2 Walks, Trails, Paths, Circuits, Connectivity, Components
10 1.3 Graph Operations
14 1.4 Cuts
18 1.5 Labeled Graphs and Isomorphism
20 II TREES
20 2.1 Trees and Forests
23 2.2 (Fundamental) Circuits and (Fundamental) Cut Sets
27 III DIRECTED GRAPHS
27 3.1 Denition
29 3.2 Directed Trees
32 3.3 Acyclic Directed Graphs
34 IV MATRICES AND VECTOR SPACES OF GRAPHS
34 4.1 Matrix Representation of Graphs
36 4.2 Cut Matrix
40 4.3 Circuit Matrix
43 4.4 An Application: Stationary Linear Networks
48 4.5 Matrices over GF(2) and Vector Spaces of Graphs
50 V GRAPH ALGORITHMS
50 5.1 Computational Complexity of Algorithms
52 5.2 Reachability: Warshalls Algorithm
53 5.3 Depth-First and Breadth-First Searches
61 5.4 The Lightest Path: Dijkstras Algorithm
63 5.5 The Lightest Path: Floyds Algorithm
66 5.6 The Lightest Spanning Tree: Kruskals and Prims Algorithms
71 5.7 The Lightest Hamiltonian Circuit (Travelling Salesmans Problem): The Annealing
Algorithm and the KarpHeld Heuristics
76 5.8 Maximum Matching in Bipartite Graphs: The Hungarian Algorithm
80 5.9 Maximum Flow in a Transport Network: The FordFulkerson Algorithm
i
ii
85 VI DRAWING GRAPHS
85 6.1 Planarity and Planar Embedding
90 6.2 The DavidsonHarel Algorithm
92 VII MATROIDS
92 7.1 Hereditary Systems
93 7.2 The Circuit Matroid of a Graph
96 7.3 Other Basic Matroids
98 7.4 Greedy Algorithm
100 7.5 The General Matroid
102 7.6 Operations on Matroids
106 References
108 Index
Foreword
These lecture notes were translated from the Finnish lecture notes for the TUT course Graa-
teoria. The laborious bulk translation was taken care of by the students Janne Tamminen (TUT)
and Kung-Chung Lee (visiting from the University of British Columbia). Most of the material
was then checked by professor Robert Pich. I want to thank the translation team for their effort.
The notes form the base text for the course MAT-41196 Graph Theory. They contain
an introduction to basic concepts and results in graph theory, with a special emphasis put on
the network-theoretic circuit-cut dualism. In many ways a model was the elegant and careful
presentation of SWAMY & THULASIRAMAN, especially the older (and better) edition. There are
of course many modern text-books with similar contents, e.g. the popular GROSS & YELLEN.
One of the usages of graph theory is to give a unied formalism for many very different-
looking problems. It then sufces to present algorithms in this common formalism. This has
lead to the birth of a special class of algorithms, the so-called graph algorithms. Half of the
text of these notes deals with graph algorithms, again putting emphasis on network-theoretic
methods. Only basic algorithms, applicable to problems of moderate size, are treated here.
Special classes of algorithms, such as those dealing with sparse large graphs, small-world
graphs, or parallel algorithms will not be treated. In these algorithms, data structure issues have
a large role, too (see e.g. SKIENA).
The basis of graph theory is in combinatorics, and the role of graphics is only in visual-
izing things. Graph-theoretic applications and models usually involve connections to the real
world on the one handoften expressed in vivid graphical termsand the denitional and
computational methods given by the mathematical combinatoric and linear-algebraic machin-
ery on the other. For many, this interplay is what makes graph theory so interesting. There is
a part of graph theory which actually deals with graphical drawing and presentation of graphs,
briey touched in Chapter 6, where also simple algorithms are given for planarity testing and
drawing. The presentation of the matter is quite supercial, a more profound treatment would
require some rather deep results in topology and curve theory. Chapter 7 contains a brief intro-
duction to matroids, a nice generalization and substitute for graphs in many ways.
Proofs of graph-theoretic results and methods are usually not given in a completely rigorous
combinatoric form, but rather using the possibilities of visualization given by graphical presen-
tations of graphs. This can lead to situations where the reader may not be completely convinced
of the validity of proofs and derivations. One of the goals of a course in graph theory must then
iii
be to provide the student with the correct touch to such seemingly loose methods of proof.
This is indeed necessary, as a completely rigoristic mathematical presentation is often almost
unreadable, whereas an excessively slack and lacunar presentation is of course useless.
Keijo Ruohonen
Chapter 1
Denitions and Fundamental Concepts
1.1 Denitions
Conceptually, a graph is formed by vertices and edges connecting the vertices.
Example.
Formally, a graph is a pair of sets (V, E), where V is the set of vertices and E is the set of
edges, formed by pairs of vertices. E is a multiset, in other words, its elements can occur more
than once so that every element has a multiplicity. Often, we label the vertices with letters (for
example: a, b, c, . . . or v
1
, v
2
, . . . ) or numbers 1, 2, . . . Throughout this lecture material, we will
label the elements of V in this way.
Example. (Continuing from the previous example) We label the vertices as follows:
v
2 v
3
v
1
v
4
v
5
We have V = {v
1
, . . . , v
5
} for the vertices and E = {(v
1
, v
2
), (v
2
, v
5
), (v
5
, v
5
), (v
5
, v
4
), (v
5
, v
4
)}
for the edges.
Similarly, we often label the edges with letters (for example: a, b, c, . . . or e
1
, e
2
, . . . ) or num-
bers 1, 2, . . . for simplicity.
1
CHAPTER 1. DEFINITIONS AND FUNDAMENTAL CONCEPTS 2
Remark. The two edges (u, v) and (v, u) are the same. In other words, the pair is not ordered.
Example. (Continuing from the previous example) We label the edges as follows:
v
2 v
3
v
1
v
4
v
5
e
1
e
2
e
3
e
4
e
5
So E = {e
1
, . . . , e
5
}.
We have the following terminologies:
1. The two vertices u and v are end vertices of the edge (u, v).
2. Edges that have the same end vertices are parallel.
3. An edge of the form (v, v) is a loop.
4. A graph is simple if it has no parallel edges or loops.
5. A graph with no edges (i.e. E is empty) is empty.
6. A graph with no vertices (i.e. V and E are empty) is a null graph.
7. A graph with only one vertex is trivial.
8. Edges are adjacent if they share a common end vertex.
9. Two vertices u and v are adjacent if they are connected by an edge, in other words, (u, v)
is an edge.
10. The degree of the vertex v, written as d(v), is the number of edges with v as an end vertex.
By convention, we count a loop twice and parallel edges contribute separately.
11. A pendant vertex is a vertex whose degree is 1.
12. An edge that has a pendant vertex as an end vertex is a pendant edge.
13. An isolated vertex is a vertex whose degree is 0.
Example. (Continuing from the previous example)
v
4
and v
5
are end vertices of e
5
.
e
4
and e
5
are parallel.
e
3
is a loop.
The graph is not simple.
e
1
and e
2
are adjacent.
v
1
and v
2
are adjacent.
The degree of v
1
is 1 so it is a pendant vertex.
e
1
is a pendant edge.
The degree of v
5
is 5.
The degree of v
4
is 2.
The degree of v
3
is 0 so it is an isolated vertex.
In the future, we will label graphs with letters, for example:
G = (V, E).
The minimum degree of the vertices in a graph G is denoted (G) (= 0 if there is an isolated
vertex in G). Similarly, we write (G) as the maximum degree of vertices in G.
Example. (Continuing from the previous example) (G) = 0 and (G) = 5.
Remark. In this course, we only consider nite graphs, i.e. V and E are nite sets.
Since every edge has two end vertices, we get
Theorem 1.1. The graph G = (V, E), where V = {v
1
, . . . , v
n
} and E = {e
1
, . . . , e
m
}, satises
n
i=1
d(v
i
) = 2m.
Corollary. Every graph has an even number of vertices of odd degree.
Proof. If the vertices v
1
, . . . , v
k
have odd degrees and the vertices v
k+1
, . . . , v
n
have even de-
grees, then (Theorem 1.1)
d(v
1
) + + d(v
k
) = 2md(v
k+1
) d(v
n
)
is even. Therefore, k is even.
Example. (Continuing from the previous example) Now the sum of the degrees is 1 + 2 + 0 +
2 + 5 = 10 = 2 5. There are two vertices of odd degree, namely v
1
and v
5
.
Asimple graph that contains every possible edge between all the vertices is called a complete
graph. A complete graph with n vertices is denoted as K
n
. The rst four complete graphs are
given as examples:
K
1
K
2
K
3
K
4
The graph G
1
= (V
1
, E
1
) is a subgraph of G
2
= (V
2
, E
2
) if
1. V
1
V
2
and
2. Every edge of G
1
is also an edge of G
2
.
Example. We have the graph
G
2
:
e
1
v
1
v
2
e
2
e
3
v
3
e
4
v
4
e
5
v
5
e
6
and some of its subgraphs are
G
1
:
e
1
v
1
v
2
G
1
:
e
1
v
1
v
2
e
2
v
3
e
4
v
4
e
5
v
5
e
6
G
1
:
v
1
v
2
v
3
e
5
v
5
e
6
and
G
1
:
v
5
e
6
The subgraph of G = (V, E) induced by the edge set E
1
E is:
G
1
= (V
1
, E
1
) =
def.
E
1
,
where V
1
consists of every end vertex of the edges in E
1
.
Example. (Continuing from above) From the original graph G, the edges e
2
, e
3
and e
5
induce
the subgraph
e
2
,e
3
,e
5
:
v
1
v
2
e
2
e
3
v
3
e
5
v
5
The subgraph of G = (V, E) induced by the vertex set V
1
V is:
G
1
= (V
1
, E
1
) =
def.
V
1
,
where E
1
consists of every edge between the vertices in V
1
.
Example. (Continuing from the previous example) From the original graph G, the vertices v
1
,
v
3
and v
5
induce the subgraph
v
1
e
3
v
3
e
5
v
5
e
6
v
1
,v
3
,v
5
:
A complete subgraph of G is called a clique of G.
1.2 Walks, Trails, Paths, Circuits, Connectivity, Components
Remark. There are many different variations of the following terminologies. We will adhere to
the denitions given here.
A walk in the graph G = (V, E) is a nite sequence of the form
v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
,
which consists of alternating vertices and edges of G. The walk starts at a vertex. Vertices v
i
t1
and v
it
are end vertices of e
jt
(t = 1, . . . , k). v
i
0
is the initial vertex and v
i
k
is the terminal
vertex. k is the length of the walk. A zero length walk is just a single vertex v
i
0
. It is allowed to
visit a vertex or go through an edge more than once. A walk is open if v
i
0
= v
i
k
. Otherwise it
is closed.
Example. In the graph
v
6
G:
v
1
e
10
e
9
e
8
e
1
v
2
e
7
e
2
v
5
e
6
e
5
v
4
v
3
e
3
e
4
the walk
v
2
, e
7
, v
5
, e
8
, v
1
, e
8
, v
5
, e
6
, v
4
, e
5
, v
4
, e
5
, v
4
is open. On the other hand, the walk
v
4
, e
5
, v
4
, e
3
, v
3
, e
2
, v
2
, e
7
, v
5
, e
6
, v
4
is closed.
A walk is a trail if any edge is traversed at most once. Then, the number of times that the
vertex pair u, v can appear as consecutive vertices in a trail is at most the number of parallel
edges connecting u and v.
Example. (Continuing from the previous example) The walk in the graph
v
1
, e
8
, v
5
, e
9
, v
1
, e
1
, v
2
, e
7
, v
5
, e
6
, v
4
, e
5
, v
4
, e
4
, v
4
is a trail.
A trail is a path if any vertex is visited at most once except possibly the initial and terminal
vertices when they are the same. A closed path is a circuit. For simplicity, we will assume in
the future that a circuit is not empty, i.e. its length 1. We identify the paths and circuits with
the subgraphs induced by their edges.
Example. (Continuing from the previous example) The walk
v
2
, e
7
, v
5
, e
6
, v
4
, e
3
, v
3
is a path and the walk
v
2
, e
7
, v
5
, e
6
, v
4
, e
3
, v
3
, e
2
, v
2
is a circuit.
The walk starting at u and ending at v is called an uv walk. u and v are connected if there
is a uv walk in the graph (then there is also a uv path!). If u and v are connected and v and w
are connected, then u and w are also connected, i.e. if there is a uv walk and a vw walk, then
there is also a uw walk. A graph is connected if all the vertices are connected to each other.
(A trivial graph is connected by convention.)
Example. The graph
is not connected.
The subgraph G
1
(not a null graph) of the graph G is a component of G if
1. G
1
is connected and
2. Either G
1
is trivial (one single isolated vertex of G) or G
1
is not trivial and G
1
is the
subgraph induced by those edges of G that have one end vertex in G
1
.
Different components of the same graph do not have any common vertices because of the fol-
lowing theorem.
Theorem 1.2. If the graph G has a vertex v that is connected to a vertex of the component G
1
of G, then v is also a vertex of G
1
.
Proof. If v is connected to vertex v
of G
1
, then there is a walk in G
v = v
i
0
, e
j
1
, v
i
1
, . . . , v
i
k1
, e
j
k
, v
i
k
= v
.
Since v
is a vertex of G
1
, then (condition #2 above) e
j
k
is an edge of G
1
and v
i
k1
is a vertex
of G
1
. We continue this process and see that v is a vertex of G
1
.
Example.
G:
v
1
v
3
v
2
e
1
e
2
v
4
e
3
e
5
v
6
e
4
v
5
e
6
v
7
e
7
v
8
G
1
G
2
G
3
G
4
The components of G are G
1
, G
2
, G
3
and G
4
.
Theorem 1.3. Every vertex of G belongs to exactly one component of G. Similarly, every edge
of G belongs to exactly one component of G.
Proof. We choose a vertex v in G. We do the following as many times as possible starting with
V
1
= {v}:
() If v
is a vertex of G such that v
/ V
1
and v
is connected to some vertex of V

1
, then
V
1
V
1
{v
}.
Since there is a nite number of vertices in G, the process stops eventually. The last V
1
induces a
subgraph G
1
of G that is the component of G containing v. G
1
is connected because its vertices
are connected to v so they are also connected to each other. Condition #2 holds because we can
not repeat (). By Theorem 1.2, v does not belong to any other component.
The edges of the graph are incident to the end vertices of the components.
Theorem 1.3 divides a graph into distinct components. The proof of the theorem gives an
algorithm to do that. We have to repeat what we did in the proof as long as we have free
vertices that do not belong to any component. Every isolated vertex forms its own component.
A connected graph has only one component, namely, itself.
A graph G with n vertices, m edges and k components has the rank
(G) = n k.
The nullity of the graph is
(G) = mn + k.
We see that (G) 0 and (G) + (G) = m. In addition, (G) 0 because
Theorem 1.4. (G) m
Proof. We will use the second principle of induction (strong induction) for m.
Induction Basis: m = 0. The components are trivial and n = k.
Induction Hypothesis: The theorem is true for m < p. (p 1)
Induction Statement: The theorem is true for m = p.
Induction Statement Proof: We choose a component G
1
of G which has at least one edge.
We label that edge e and the end vertices u and v. We also label G
2
as the subgraph of G and
G
1
, obtained by removing the edge e from G
1
(but not the vertices u and v). We label G
as the
graph obtained by removing the edge e from G (but not the vertices u and v) and let k
be the
number of components of G
. We have two cases:

1. G
2
is connected. Then, k
= k. We use the Induction Hypothesis on G
:
n k = n k
= (G
) m1 < m.
2. G
2
is not connected. Then there is only one path between u and v:
u, e, v
and no other path. Thus, there are two components in G
2
and k
= k + 1. We use the
Induction Hypothesis on G
:
(G
) = n k
= n k 1 m1.
Hence n k m.
These kind of combinatorial results have many consequences. For example:
Theorem 1.5. If G is a connected graph and k 2 is the maximum path length, then any two
paths in G with length k share at least one common vertex.
Proof. We only consider the case where the paths are not circuits (Other cases can be proven in
a similar way.). Consider two paths of G with length k:
v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
(path p
1
)
and
v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
(path p
2
).
Let us consider the counter hypothesis: The paths p
1
and p
2
do not share a common vertex.
Since G is connected, there exists an v
i
0
v
i
k
path. We then nd the last vertex on this path
which is also on p
1
(at least v
i
0
is on p
1
) and we label that vertex v
it
. We nd the rst vertex of
the v
it
v
i
k
path which is also on p
2
(at least v
i
k
is on p
2
) and we label that vertex v
i
s
. So we get
a v
it
v
i
s
path
v
it
, e
j
1
, . . . , e
j
, v
i
s
.
The situation is as follows:
v
i
0
, e
j
1
, v
i
1
, . . . ,v
it
, e
j
t+1
, . . . , e
j
k
, v
i
k
e
j
1
.
.
.
e
j
v
i
0
, e
j
1
, v
i
1
, . . . ,v
i
s
, e
j
s+1
, . . . , e
j
k
, v
i
k
From here we get two paths: v
i
0
v
i
k
path and v
i
0
v
i
k
path. The two cases are:
t s: Now the length of the v
i
0
v
i
k
path is k + 1.
1
t < s: Now the length of the v
i
0
v
i
k
path is k + 1.
A graph is circuitless if it does not have any circuit in it.

Theorem 1.6. A graph is circuitless exactly when there are no loops and there is at most one
path between any two given vertices.
Proof. First let us assume G is circuitless. Then, there are no loops in G. Let us assume the
counter hypothesis: There are two different paths between distinct vertices u and v in G:
u = v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
= v (path p
1
)
and
u = v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
, v
i
= v (path p
2
)
(here we have i
0
= i
0
and i
k
= i
), where k . We choose the smallest index t such that

v
it
= v
i
t
.
There is such a t because otherwise
1
From now on, the symbol
means contradiction. If we get a contradiction by proceeding from the assump-

tions, the hypothesis must be wrong.
1. k > and v
i
k
= v = v
i
= v
i
) or
2. k = and v
i
0
= v
i
0
, . . . , v
i
= v
i
. Then, there would be two parallel edges between two

consecutive vertices in the path. That would imply the existence of a circuit between two
vertices in G.
u
v
v
v
i
t1
i
s
p
1
p
2
We choose the smallest index s such that s t and v
is
is in the path p
2
(at least v
i
k
is in p
2
). We
choose an index r such that r t and v
i
r
= v
is
(it exists because p
1
is a path). Then,
v
i
t1
, e
jt
, . . . , e
js
, v
is
(= v
i
r
), e
j
r
, . . . , e
j
t
, v
i
t1
(= v
i
t1
)
is a circuit.
(Verify the case t = s = r.)

Let us prove the reverse implication. If the graph does not have any loops and no two distinct
vertices have two different paths between them, then there is no circuit. For example, if
v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
= v
i
0
is a circuit, then either k = 1 and e
j
1
is a loop (
), or k 2 and the two vertices v

i
0
and v
i
1
are connected by two distinct paths
v
i
0
, e
j
1
, v
i
1
and v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
= v
i
0
(
).
1.3 Graph Operations
The complement of the simple graph G = (V, E) is the simple graph G = (V, E), where the
edges in E are exactly the edges not in G.
Example.
v
2
v
1
v
3
v
4
v
5
G:
v
2
v
1
v
3
v
4
v
5
G:
_
Example. The complement of the complete graph K
n
is the empty graph with n vertices.
Obviously, G = G. If the graphs G = (V, E) and G
= (V

, E
) are simple and V

V , then
the difference graph is G G
= (V, E
), where E
contains those edges from G that are not

in G
(simple graph).
Example.
G: G':
G G':
Here are some binary operations between two simple graphs G
1
= (V
1
, E
1
) and G
2
=
(V
2
, E
2
):
The union is G
1
G
2
= (V
1
V
2
, E
1
E
2
) (simple graph).
The intersection is G
1
G
2
= (V
1
V
2
, E
1
E
2
) (simple graph).
The ring sumG
1
G
2
is the subgraph of G
1
G
2
induced by the edge set E
1
E
2
(simple
graph). Note! The set operation is the symmetric difference, i.e.
E
1
E
2
= (E
1
E
2
) (E
2
E
1
).
Since the ring sum is a subgraph induced by an edge set, there are no isolated vertices. All three
operations are commutative and associative.
Example. For the graphs
G
1
: G
2
:
v
1
v
2
v
5
v
3
v
4
v
1
v
3
v
6
v
7
e
1
e
2
e
3
e
5
e
4
e
1
e
7
e
6
we have
v
6
v
7
e
7
e
6
v
1
v
2
v
5
v
3
v
4
e
1
e
2
e
3
e
5
e
4
G
1
G
2
:
G
1
G
2
:
v
1
v
3
e
1
v
6
e
6
G
1
G
2
:
v
1
v
2
v
3
v
4
e
2
e
3
e
5
e
4
e
7
Remark. The operations , and can also be dened for more general graphs other than
simple graphs. Naturally, we have to keep track of the multiplicity of the edges:
: The multiplicity of an edge in G
1
G
2
is the larger of its multiplicities in G
1
and G
2
.
1
G
2
is the smaller of its multiplicities in G
1
and G
2
.
1
G
2
is |m
1
m
2
|, where m
1
is its multiplicity in G
1
and m
2
is its multiplicity in G
2
.
(We assume zero multiplicity for the absence of an edge.) In addition, we can generalize the dif-
ference operation for all kinds of graphs if we take account of the multiplicity. The multiplicity
of the edge e in the difference GG
is
m
1

m
2
=
m
1
m
2
, if m
1
m
2
0, if m
1
< m
2
(also known as the proper difference),
where m
1
and m
2
are the multiplicities of e in G
1
and G
2
, respectively.
If v is a vertex of the graph G = (V, E), then G v is the subgraph of G induced by the
vertex set V {v}. We call this operation the removal of a vertex.
G
1
v
4
:
v
1
v
5
v
3
e
1
e
3
v
2
Similarly, if e is an edge of the graph G = (V, E), then G e is graph (V, E
), where E
is
obtained by removing e from E. This operation is known as removal of an edge. We remark
that we are not talking about removing an edge as in Set Theory, because the edge can have
nonunit multiplicity and we only remove the edge once.
G
1
e
5
:
v
1
v
2
v
5
v
3
v
4
e
1
e
2
e
3
e
4
If u and v are two distinct vertices of the graph G = (V, E), then we can short-circuit the
two vertices u and v and obtain the graph (V

, E
), where
V

= (V {u, v}) {w} (w / V is the new vertex)
and
E
= (E {(v
, u), (v
, v) | v
V }) {(v
, w) | (v
, u) E or (v
, v) E}
{(w, w) | (u, u) E or (v, v) E}
(Recall that the pair of vertices corresponding to an edge is not ordered). Note! We have to
maintain the multiplicity of the edges. In particular, the edge (u, v) becomes a loop.
Example. (Continuing from the previous example) Short-circuit v
3
and v
4
in the graph G
1
:
v
1
v
2
v
5
w
In the graph G = (V, E), contracting the edge e = (u, v) (not a loop) means the operation in
which we rst remove e and then short-circuit u and v. (Contracting a loop simply removes that
loop.)
Example. (Continuing from the previous example) We contract the edge e
3
in G
1
by rst re-
moving e
3
and then short-circuiting v
2
and v
3
.
v
1
v
2
v
5
v
3
v
4
e
1
e
2
e
5
e
4
v
1
v
5
w v
4
e
2
Remark. If we restrict short-circuiting and contracting to simple graphs, then we remove loops
and all but one of the parallel edges between end vertices from the results.
1.4 Cuts
A vertex v of a graph G is a cut vertex or an articulation vertex of G if the graph Gv consists
of a greater number of components than G.
Example. v is a cut vertex of the graph below:
cut vertex
G:
G v:
v
(Note! Generally, the only vertex of a trivial graph is not a cut vertex, neither is an isolated
vertex.)
A graph is separable if it is not connected or if there exists at least one cut vertex in the
graph. Otherwise, the graph is nonseparable.
Example. The graph G in the previous example is separable.
Example. The graph below is nonseparable.
A block of the graph G is a subgraph G
1
of G (not a null graph) such that
G
1
is nonseparable, and
if G
2
is any other subgraph of G, then G
1
G
2
= G
1
or G
1
G
2
is separable (think about
that!).
Example. The graph below is separable:
cut vertex
Theorem 1.7. The vertex v is a cut vertex of the connected graph G if and only if there exist
two vertices u and w in the graph G such that
(i) v = u, v = w and u = w, but
(ii) v is on every uw path.
Proof. First, let us consider the case that v is a cut-vertex of G. Then, G v is not connected
and there are at least two components G
1
= (V
1
, E
1
) and G
2
= (V
2
, E
2
). We choose u V
1
and w V
2
. The uw path is in G because it is connected. If v is not on this path, then the path
is also in Gv (
). The same reasoning can be used for all the uw paths in G.

If v is in every uw path, then the vertices u and w are not connected in Gv.
Theorem 1.8. A nontrivial simple graph has at least two vertices which are not cut vertices.
Proof. We will use induction for the graph G with n vertices.
Induction Basis: The case n = 2 is obviously true.
Induction Hypothesis: The theorem is true for n k. (k 2)
Induction Statement: The theorem is true for n = k + 1.
Induction Statement Proof: If there are no cut vertices in G, then it is obvious. Otherwise,
we consider a cut vertex v of G. Let G
1
, . . . , G
m
be the components of G v (so m 2).
Every component G
i
falls into one of the two cases:
1. G
i
is trivial so the only vertex of G
i
is a pendant vertex or an isolated vertex of G but it is
not a cut vertex of G.
2. G
i
is not trivial. The Induction Hypothesis tells us that there exist two vertices u and w
in G
i
which are not cut vertices of G
i
. If v and u (respectively v and w) are not adjacent
in G, then u (respectively w) is not a cut vertex in G. If both v and u as well as u and w
are adjacent in G, then u and w can not be cut vertices of G.
A cut set of the connected graph G = (V, E) is an edge set F E such that
1. GF (remove the edges of F one by one) is not connected, and
2. GH is connected whenever H F.
Theorem 1.9. If F is a cut set of the connected graph G, then GF has two components.
Proof. Let F = {e
1
, . . . , e
k
}. The graph G{e
1
, . . . , e
k1
} is connected (and so is G if k = 1)
by condition #2. When we remove the edges from the connected graph, we get at most two
components.
Example. In the graph
v
1
e
1
v
2
e
4
v
3
e
3
e
2
e
5
e
6
v
4
e
7
e
8
v
6
v
5
{e
1
, e
4
}, {e
6
, e
7
}, {e
1
, e
2
, e
3
}, {e
8
}, {e
3
, e
4
, e
5
, e
6
}, {e
2
, e
5
, e
7
}, {e
2
, e
5
, e
6
} and {e
2
, e
3
, e
4
} are
cut sets. Are there other cut sets?
In a graph G = (V, E), a pair of subsets V
1
and V
2
of V satisfying
V = V
1
V
2
, V
1
V
2
= , V
1
= , V
2
= ,
is called a cut (or a partition) of G, denoted V
1
, V
2
. Usually, the cuts V
1
, V
2
and V
2
, V
1
are
considered to be the same.
Example. (Continuing from the previous example) {v
1
, v
2
, v
3
}, {v
4
, v
5
, v
6
} is a cut.
We can also think of a cut as an edge set:
cut V
1
, V
2
= {those edges with one end vertex in V
1
and the other end vertex in V
2
}.
(Note! This edge set does not dene V
1
and V
2
uniquely so we can not use this for the denition
of a cut.)
Using the previous denitions and concepts, we can easily prove the following:
1. The cut V
1
, V
2
of a connected graph G (considered as an edge set) is a cut set if and
only if the subgraphs induced by V
1
and V
2
are connected, i.e. G V
1
, V
2
has two
components.
2. If F is a cut set of the connected graph G and V
1
and V
2
are the vertex sets of the two
components of GF, then V
1
, V
2
is a cut and F = V
1
, V
2
.
3. If v is a vertex of a connected (nontrivial) graph G = (V, E), then {v}, V {v} is a cut
of G. It follows that the cut is a cut set if the subgraph (i.e. G v) induced by V {v}
is connected, i.e. if v is not a cut vertex.
If there exists a cut V
1
, V
2
for the graph G = (V, E) so that E = V
1
, V
2
, i.e. the cut
(considered as an edge set) includes every edge, then the graph G is bipartite.
Example. The graph
v
1
v
2
v
3
v
4
v
5
v
6
v
7
is bipartite. V
1
= {v
1
, v
2
, v
3
} and V
2
= {v
4
, v
5
, v
6
, v
7
}.
A simple bipartite graph is called a complete bipartite graph if we can not possibly add any
more edges to the edge set V
1
, V
2
, i.e. the graph contains exactly all edges that have one end
vertex in V
1
and the other end vertex in V
2
. If there are n vertices in V
1
and m vertices in V
2
, we
denote it as K
n,m
(cf. complete graph).
Example.
K
1,1
: K
1,2
:
K
2,1
:
K
2,3
:
(Usually K
n,m
and K
m,n
are considered to be the same.)
1.5 Labeled Graphs and Isomorphism
By a labeling of the vertices of the graph G = (V, E), we mean a mapping : V A, where
A is called the label set. Similarly, a labeling of the edges is a mapping : E B, where B is
the label set. Often, these labels are numbers. Then, we call them weights of vertices and edges.
In a weighted graph, the weight of a path is the sum of the weights of the edges traversed.
The labeling of the vertices (respectively edges) is injective if distinct vertices (respectively
edges) have distinct labels. An injective labeling is bijective if there are as many labels in A
(respectively in B) as the number of vertices (respectively edges).
Example. If A = {0, 1} and B = R, then in the graph,
0
1
1
0
0
1
0.1
0.4
0.7
0.6
0.2
the labeling of the edges (weights) is injective but not the labeling of the vertices.
The two graphs G
1
= (V
1
, E
1
) and G
2
= (V
2
, E
2
) are isomorphic if labeling the vertices of
G
1
bijectively with the elements of V
2
gives G
2
. (Note! We have to maintain the multiplicity of
the edges.)
Example. The graphs G
1
and G
2
are isomorphic and the vertex labeling v
i
v
i
and edge
labeling e
j
e
j
dene the isomorphism.
G
1
:
v
1
v
2
v
3
e
1
e
5
v
4
v
5
e
3
e
6
e
2
v
7
e
9
e
7
v
6
v
8
e
8
e
10
e
4
G
2
:
v'
v'
e'
e'
e'
v' v'
v'
e'
v'
e' e' e'
e'
v'
v'
e'
8
8
9
e'
10
6
7
5
5
4
4
2
7
2
1
3
1
3 6
Determining whether or not two graphs are isomorphic is a well researched problem. It
differs signicantly from other problems in graph theory and network analysis. In addition,
it has a lot to do with group theory in algebra. The problem is important in the theory of
Computational Complexity. For example, refer to KBLER, J. & SCHNING, U. & TORN,
J.: The Graph Isomorphism Problem. Its Structural Complexity. Birkhuser (1993).
Chapter 2
Trees
2.1 Trees and Forests
A forest is a circuitless graph. A tree is a connected forest. A subforest is a subgraph of a forest.
A connected subgraph of a tree is a subtree. Generally speaking, a subforest (respectively
subtree) of a graph is its subgraph, which is also a forest (respectively tree).
Example. Four trees which together form a forest:
A spanning tree of a connected graph is a subtree that includes all the vertices of that graph. If
T is a spanning tree of the graph G, then
GT =
def.
T
is the cospanning tree.

Example.
G: spanning tree:
20
CHAPTER 2. TREES 21
cospanning tree
The edges of a spanning tree are called branches and the edges of the corresponding cospanning
tree are called links or chords.
Theorem 2.1. If the graph G has n vertices and m edges, then the following statements are
equivalent:
(i) G is a tree.
(ii) There is exactly one path between any two vertices in G and G has no loops.
(iii) G is connected and m = n 1.
(iv) G is circuitless and m = n 1.
(v) Gis circuitless and if we add any new edge to G, then we will get one and only one circuit.
Proof. (i)(ii): If G is a tree, then it is connected and circuitless. Thus, there are no loops in
G. There exists a path between any two vertices of G. By Theorem 1.6, we know that there is
only one such path.
(ii)(iii): G is connected. Let us use induction on m.
Induction Basis: m = 0, G is trivial and the statement is obvious.
Induction Hypothesis: m = n 1 when m . ( 0)
Induction Statement: m = n 1 when m = + 1.
Induction Statement Proof: Let e be an edge in G. Then G e has edges. If G e is
connected, then there exist two different paths between the end vertices of e so (ii) is false.
Therefore, Ge has two components G
1
and G
2
. Let there be n
1
vertices and m
1
edges in G
1
.
Similarly, let there be n
2
vertices and m
2
vertices in G
2
. Then,
n = n
1
+n
2
and m = m
1
+m
2
+ 1.
The Induction Hypothesis states that
m
1
= n
1
1 and m
2
= n
2
1,
so m = n
1
+n
2
1 = n 1.
(iii)(iv): Consider the counter hypothesis: There is a circuit in G. Let e be some edge in
that circuit. Thus, there are n vertices and n 2 edges in the connected graph Ge.
1
(iv)(v): If G is circuitless, then there is at most one path between any two vertices (The-
orem 1.6). If G has more than one component, then we will not get a circuit when we draw an
edge between two different components. By adding edges, we can connect components without
creating circuits:
1
In a connected graph with n vertices, there are at least n 1 edges. (Theorem 1.4)
CHAPTER 2. TREES 22
1st component 2nd component
3rd
component
4th
component
If we add k( 1) edges, then (because (i)(iii))
m +k = n 1 (
because m = n 1).
So G is connected. When we add an edge between vertices that are not adjacent, we get only
one circuit. Otherwise, we can remove an edge from one circuit so that other circuits will not
be affected and the graph stays connected, in contradiction to (iii)(iv). Similarly, if we add a
parallel edge or a loop, we get exactly one circuit.
(v)(i): Consider the counter hypothesis: G is not a tree, i.e. it is not connected. When we
add edges as we did previously, we do not create any circuits (see gure).
Since spanning trees are trees, Theorem 2.1 is also true for spanning trees.
Theorem 2.2. A connected graph has at least one spanning tree.
Proof. Consider the connected graph G with n vertices and m edges. If m = n 1, then G
is a tree. Since G is connected, m n 1 (Theorem 1.4). We still have to consider the case
m n, where there is a circuit in G. We remove an edge e from that circuit. G e is now
connected. We repeat until there are n 1 edges. Then, we are left with a tree.
Remark. We can get a spanning tree of a connected graph by starting from an arbitrary sub-
forest M (as we did previously). Since there is no circuit whose edges are all in M, we can
remove those edges from the circuit which are not in M.
By Theorem 2.1, the subgraph G
1
of G with n vertices is a spanning tree of G (thus G is
connected) if any three of the following four conditions hold:
1. G
1
has n vertices.
2. G
1
is connected.
3. G
1
has n 1 edges.
4. G
1
is circuitless.
Actually, conditions #3 and #4 are enough to guarantee that G
1
is a spanning tree. If conditions
#3 and #4 hold but G
1
is not connected, then the components of G
1
are trees and the number of
edges in G
1
would be
number of vertices number of components < n 1 (
).
CHAPTER 2. TREES 23
Theorem 2.3. If a tree is not trivial, then there are at least two pendant vertices.
Proof. If a tree has n( 2) vertices, then the sum of the degrees is 2(n 1). If every vertex
has a degree 2, then the sum will be 2n (
). On the other hand, if all but one vertex have

degree 2, then the sum would be 1 + 2(n 1) = 2n 1 (
). (This also follows from

Theorem 1.8 because a cut vertex of a tree is not a pendant vertex!)
A forest with k components is sometimes called a k-tree. (So a 1-tree is a tree.)
Example.
4-tree:
We use Theorem 2.1 to see that a graph with k components has a spanning k-tree, also known
as a spanning forest, which has k components.
2.2 (Fundamental) Circuits and (Fundamental) Cut Sets
If the branches of the spanning tree T of a connected graph G are b
1
, . . . , b
n1
and the corre-
sponding links of the cospanning tree T
are c
1
, . . . , c
mn+1
, then there exists one and only one
circuit C
i
in T + c
i
(which is the subgraph of G induced by the branches of T and c
i
) (The-
orem 2.1). We call this circuit a fundamental circuit. Every spanning tree denes m n + 1
fundamental circuits C
1
, . . . , C
mn+1
, which together form a fundamental set of circuits. Every
fundamental circuit has exactly one link which is not in any other fundamental circuit in the
fundamental set of circuits. Therefore, we can not write any fundamental circuit as a ring sum
of other fundamental circuits in the same set. In other words, the fundamental set of circuits is
linearly independent under the ring sum operation.
Example.
G:
T: T*:
c
1
c
2
c
3
CHAPTER 2. TREES 24
C
1
C
2
C
3
The graph T b
i
has two components T
1
and T
2
. The corresponding vertex sets are V
1
and
V
2
. Then, V
1
, V
2
is a cut of G. It is also a cut set of G if we treat it as an edge set because
G V
1
, V
2
has two components (result #1 p. 17). Thus, every branch b
i
of T has a corre-
sponding cut set I
i
. The cut sets I
1
, . . . , I
n1
are also known as fundamental cut sets and they
form a fundamental set of cut sets. Every fundamental cut set includes exactly one branch of T
and every branch of T belongs to exactly one fundamental cut set. Therefore, every spanning
tree denes a unique fundamental set of cut sets for G.
Example. (Continuing from the previous example) The graph
G:
e
2
e
1
e
4
e
3
e
5
e
6
e
7
e
8
has the spanning tree
T:
b
1
b
2
b
3
b
4
b
5
that denes these fundamental cut sets:
b
1
: e
1
, e
2
b
2
: e
2
, e
3
, e
4
b
3
: e
2
, e
4
, e
5
, e
6
b
4
: e
2
, e
4
, e
5
, e
7
b
5
: e
8
Next, we consider some properties of circuits and cut sets:

(a) Every cut set of a connected graph G includes at least one branch from every spanning
tree of G. (Counter hypothesis: Some cut set F of G does not include any branches of a
spanning tree T. Then, T is a subgraph of GF and GF is connected.
)
(b) Every circuit of a connected graph Gincludes at least one link from every cospanning tree
of G. (Counter hypothesis: Some circuit C of Gdoes not include any link of a cospanning
tree T
. Then, T = GT
has a circuit and T is not a tree.
)
CHAPTER 2. TREES 25
Theorem 2.4. The edge set F of the connected graph G is a cut set of G if and only if
(i) F includes at least one branch from every spanning tree of G, and
(ii) if H F, then there is a spanning tree none of whose branches is in H.
Proof. Let us rst consider the case where F is a cut set. Then, (i) is true (previous proposition
(a)). If H F then G H is connected and has a spanning tree T. This T is also a spanning
tree of G. Hence, (ii) is true.
Let us next consider the case where both (i) and (ii) are true. Then G F is disconnected.
If H F there is a spanning tree T none of whose branches is in H. Thus T is a subgraph of
GH and GH is connected. Hence, F is a cut set.
Similarly:
Theorem 2.5. The subgraph C of the connected graph G is a circuit if and only if
(i) C includes at least one link from every cospanning tree of G, and
(ii) if D is a subgraph of C and D ,= C, then there exists a cospanning tree none of whose
links is in D.
Proof. Let us rst consider the case where C is a circuit. Then, C includes at least one link
from every cospanning tree (property (b) above) so (i) is true. If D is a proper subgraph of C,
it obviously does not contain circuits, i.e. it is a forest. We can then supplement D so that it is
a spanning tree of G (see remark on p. 22), i.e. some spanning tree T of G includes D and D
does not include any link of T
. Thus, (ii) is true.

Now we consider the case where (i) and (ii) are both true. Then, there has to be at least one
circuit in C because C is otherwise a forest and we can supplement it so that it is a spanning
tree of G (see remark on p. 22). We take a circuit C
in C. Since (ii) is true, C
,= C is not
true, because C
is a circuit and it includes a link from every cospanning tree (see property (b)
above). Therefore, C = C
is a circuit.
Theorem 2.6. A circuit and a cut set of a connected graph have an even number of common
edges.
Proof. We choose a circuit C and a cut set F of the connected graph G. G F has two
components G
1
= (V
1
, E
1
) and G
2
= (V
2
, E
2
). If C is a subgraph of G
1
or G
2
, then the
theorem is obvious because they have no common edges. Let us assume that C and F have
common edges. We traverse around a circuit by starting at some vertex v of G
1
. Since we come
back to v, there has to be an even number of edges of the cut V
1
, V
2
in C.
The reader is advised to read the following several times:
Theorem 2.7. A fundamental circuit corresponding to link c of the cospanning tree T
of a
connected graph is formed exactly by those branches of T whose corresponding fundamental
cut set includes c.
Proof. There exists a fundamental circuit C that corresponds to link c of T
. The other edges

b
1
, . . . , b
k
of C are branches of T. We denote I
i
as the fundamental cut set that corresponds to
branch b
i
. Then, b
i
is the only branch of T which is in both C and I
i
. On the other hand, c is
the only link of T
in C. By Theorem 2.6, we know that the common edges of C and I

i
are b
i
and c, in other words, c is an edge of I
i
. Then, we show that there is no c in the fundamental cut
sets I
k+1
, . . . , I
n1
that correspond to the branches b
k+1
, . . . , b
n1
of T. For instance, if c were
in I
k+1
, then the fundamental cut set I
k+1
and the circuit C would have exactly one common
edge. (
). So c is only in the fundamental cut sets I

1
, . . . , I
k
.
CHAPTER 2. TREES 26
The following is the corresponding theorem for fundamental cut sets:
Theorem 2.8. The fundamental cut set corresponding to branch b of the spanning tree T of a
connected graph consists exactly of those links of T
whose corresponding fundamental circuit

includes b.
Proof. Let I be a fundamental cut set that corresponds to the branch b of T. Other edges
c
1
, . . . , c
k
of I are links of T
. Let C
i
denote the fundamental circuit that corresponds to c
i
.
Then, c
i
is the only link of T
in both I and C
i
. On the other hand, b is the only branch of T in
I. By Theorem 2.6, the common edges of I and C
i
are b and c
i
, in other words, b is an edge of
C
i
. Then, we show that the fundamental circuits C
k+1
, . . . , C
mn+1
corresponding to the links
c
k+1
, . . . , c
mn+1
do not include b. For example, if b were in C
k+1
, then the fundamental circuit
C
k+1
and the cut set I would have exactly one common edge (
). Hence, the branch b is only

in fundamental circuits C
1
, . . . , C
k
.
From the results, we can see the duality between cut sets and circuits of a graph: The
theorems for cut sets can generally be converted to dual theorems for circuits and vice versa.
Usually, we just need to change some of the key terminologies to their duals in the theorems and
proofs. In particular, we take advantage of this dualism for dealing with matroids (see Chapter
7).
Chapter 3
Directed Graphs
3.1 Denition
Intuitively, a directed graph or digraph is formed by vertices connected by directed edges or
arcs.
1
Example.
Formally, a digraph is a pair (V, E), where V is the vertex set and E is the set of vertex pairs as
in usual graphs. The difference is that now the elements of E are ordered pairs: the arc from
vertex u to vertex v is written as (u, v) and the other pair (v, u) is the opposite direction arc. We
also have to keep track of the multiplicity of the arc (direction of a loop is irrelevant). We can
pretty much use the same notions and results for digraphs from Chapter 1. However:
1. Vertex u is the initial vertex and vertex v is the terminal vertex of the arc (u, v). We also
say that the arc is incident out of u and incident into v.
2. The out-degree of the vertex v is the number of arcs out of it (denoted d
+
(v)) and the
in-degree of v is the number of arcs going into it (denoted d
(v)).
3. In the directed walk (trail, path or circuit),
v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
v
i
is the initial vertex and v

i
1
is the terminal vertex of the arc e
j
. .
4. When we treat the graph (V, E) as a usual undirected graph, it is the underlying undirected
graph of the digraph G = (V, E), denoted G
u
.
1
This not a standard terminology. We will however call directed edges arcs in the sequel.
27
CHAPTER 3. DIRECTED GRAPHS 28
5. Digraph G is connected if G
u
is connected. The components of G are the directed sub-
graphs of G that correspond to the components of G
u
. The vertices of G are connected if
they are connected in G
u
. Other notions for undirected graphs can be used for digraphs
as well by dealing with the underlying undirected graph.
6. Vertices u and v are strongly connected if there is a directed uv path and also a directed
vu path in G.
7. Digraph G is strongly connected if every pair of vertices is strongly connected. By con-
vention, the trivial graph is strongly connected.
8. A strongly connected component H of the digraph G is a directed subgraph of G (not a
null graph) such that H is strongly connected, but if we add any vertices or arcs to it, then
it is not strongly connected anymore.
Every vertex of the digraph G belongs to one strongly connected component of G (compare to
Theorem 1.3). However, an arc does not necessarily belong to any strongly connected compo-
nent of G.
Example. For the digraph G
v
1
v
2
e
1
e
9
e
2
e
3
e
5
v
4
v
6
e
6
e
8
v
5
v
3
e
4
e
7
the strongly connected components are (v
1
, ), (v
2
, v
3
, v
4
, e
3
, e
4
, e
5
), (v
5
, ) and (v
6
, ).
The condensed graph G
c
of the digraph G is obtained by contracting all the arcs in every
strongly connected component.
Example. (Continuing from the previous example) The condensed graph is
v
1
e
9
v
6
e
8
v
5
w
3.2 Directed Trees
A directed graph is quasi-strongly connected if one of the following conditions holds for every
pair of vertices u and v:
(i) u = v or
(ii) there is a directed uv path in the digraph or
(iii) there is a directed vu path in the digraph or
(iv) there is a vertex w so that there is a directed wu path and a directed wv path.
Example. (Continuing from the previous example) The digraph G is quasi-strongly connected.
Quasi-strongly connected digraphs are connected but not necessarily strongly connected.
The vertex v of the digraph Gis a root if there is a directed path from v to every other vertex
of G.
Example. (Continuing from the previous example) The digraph G only has one root, v
1
.
Theorem 3.1. A digraph has at least one root if and only if it is quasi-strongly connected.
Proof. If there is a root in the digraph, it follows from the denition that the digraph is quasi-
strongly connected.
Let us consider a quasi-strongly connected digraph G and show that it must have at least
one root. If G is trivial, then it is obvious. Otherwise, consider the vertex set V = v
1
, . . . , v
n
of G where n 2. The following process shows that there must be a root:

1. Set P V .
2. If there is a directed uv path between two distinct vertices u and v in P, then we remove
v from P. Equivalently, we set P P v. We repeat this step as many times as
possible.
3. If there is only one vertex left in P, then it is the root. For other cases, there are at least
two distinct vertices u and v in P and there is no directed path between them in either
direction. Since G is quasi-strongly connected, from condition (iv) it follows that there
is a vertex w and a directed wu path as well as a directed wv path. Since u is in P, w
can not be in P. We remove u and v from P and add w, i.e. we set P P u, v and
P P w. Go back to step #2.
4. Repeat as many times as possible.
Every time we do this, there are fewer and fewer vertices in P. Eventually, we will get a root
because there is a directed path from some vertex in P to every vertex we removed from P.
The digraph G is a tree if G
u
is a tree. It is a directed tree if G
u
is a tree and G is quasi-
strongly connected, i.e. it has a root. A leaf of a directed tree is a vertex whose out-degree is
zero.
Example.
6

l
e
a
v
e
s
r
o
o
t
Theorem 3.2. For the digraph G with n > 1 vertices, the following are equivalent:
(i) G is a directed tree.
(ii) Gis a tree with a vertex from which there is exactly one directed path to every other vertex
of G.
(iii) G is quasi-strongly connected but G e is not quasi-strongly connected for any arc e in
G.
(iv) G is quasi-strongly connected and every vertex of G has an in-degree of 1 except one
vertex whose in-degree is zero.
(v) There are no circuits in G (i.e. not in G
u
) and every vertex of G has an in-degree of 1
except one vertex whose in-degree is zero.
(vi) G is quasi-strongly connected and there are no circuits in G (i.e. not in G
u
).
Proof. (i)(ii): If G is a directed tree, then there is a root. This implies that there is a directed
path from the root to every other vertex in G (but not more than one path since G
u
is a tree).
(ii)(iii): If (ii) is true, then G obviously is quasi-strongly connected. We will prove by
contradiction by considering the counter hypothesis: There is an arc e in G such that G e is
quasi-strongly connected. The arc e is not a loop because G is a directed tree. Let u and v be
the two different end vertices of e. There does not exist a directed uv path or a directed vu
path in Ge (otherwise G
u
would have a circuit). Therefore, there is a vertex w and a directed
wu path as well as a directed wv path. However, this leads to the existence of two directed
wu paths or two directed wv paths in G depending on the direction of the arc e. Then, there
is a circuit in the tree G
u
. (
by Theorem 1.6).
(iii)(iv): If G quasi-strongly connected, then it has a root r (Theorem 3.1) so that the in-
degrees of other vertices are 1. We start by considering the counter hypothesis: There exists
a vertex v ,= r and d
(v) > 1. Then, v is the terminal vertex of two distinct arcs (u, v) and
(w, v). If there were a loop e in G, then G e would be quasi-strongly connected (
). Thus,
u ,= v with w ,= v. Now, there are two distinct directed trails from r to v. The rst one includes
(u, v) and the second one includes (w, v). We have two possible cases:
r
u
w
v
r
u
w
v
In the digraph on the left, the paths ru and rw do not include the arcs (u, v) and (w, v).
Both G (u, v) and G (w, v) are quasi-strongly connected. In the digraph on the right,
the ru path includes the arc (w, v) or (as in the gure) the rw path includes the arc (u, v).
In either case, only one of G (u, v) and G (w, v) is quasi-strongly connected because
the root is r (Theorem 3.1). (
) We still have to show that d
(r) = 0. Let us consider the

counter hypothesis: d
(r) 1. Then, r is the terminal vertex of some arc e. However, the tree
Ge is then quasi-strongly connected since r is its root (Theorem 3.1). (
)
(iv)(v): If (iv) is true, then it is enough to show that there are no circuits in G
u
. The sum
of in-degrees of all the vertices in G is n 1 and the sum of out-degrees of all the vertices in G
is also n 1, i.e. there are n 1 arcs in G. Since G is quasi-strongly connected, it is connected
and it is a tree (Theorem 2.1). Therefore, there are no circuits in G
u
.
(v)(vi): If we assume that (v) is true, then there are n 1 arcs in G (compare to the
previous proof). By Theorem 2.1, G is a tree. We denote by r the vertex satisfying condition
(v). By Theorem 2.1, we see that there is exactly one path to any other vertex of G from r.
These paths are also directed. Otherwise, d
(r) 1 or the in-degree of some vertex on that

path is > 1 or the in-degree of some other vertex other than r on that path is zero. Hence, r is a
root and G is quasi-strongly connected (Theorem 3.1).
(vi)(i): If G is quasi-strongly connected, then it has a root (Theorem 3.1). Since G is
connected and there are no circuits in G, it is a tree.
A directed subgraph T of the digraph G is a directed spanning tree if T is a directed tree
and T includes every vertex of G.
Example.
G: T:
Theorem 3.3. A digraph has a directed spanning tree if and only if it is quasi-strongly con-
nected.
Proof. If the digraph G has a directed spanning tree T, then the root of T is also a root for G
and it is quasi-strongly connected (Theorem 3.1).
We now assume that G is quasi-strongly connected and show that it has a directed spanning
tree. If Gis a directed tree, then it is obvious. Otherwise, from Theorem 3.2, we know that there
is an arc e in G so that if we remove e, G remains quasi-strongly connected. We systematically
remove these kind of arcs until we get a directed tree. (Compare to the proof for Theorem
2.2)
3.3 Acyclic Directed Graphs
A directed graph with at least one directed circuit is said to be cyclic. A directed graph is acyclic
otherwise. Obviously, directed trees are acyclic but the reverse implication is not true.
Example. The digraph
is acyclic but it is not a directed tree.
Theorem 3.4. In an acyclic digraph, there exist at least one source (a vertex whose in-degree
is zero) and at least one sink (a vertex whose out-degree is zero).
Proof. Let G be an acyclic digraph. If G has no arcs, then it is obvious. Otherwise, let us
consider the directed path
v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
,
which has the maximum path length k. Since G is acyclic, v
i
0
,= v
i
k
. If (v, v
i
0
) is an arc, then
one of the following is true:
v ,= v
it
for every value of t = 0, . . . , k. Then,
v, (v, v
i
0
), v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
j
k
, v
i
k
is a directed path with length k + 1.
v = v
it
for some value of t. We choose the smallest such t. Then, t > 0 because there are
no loops in G and
v
i
0
, e
j
1
, v
i
1
, e
j
2
, . . . , e
jt
, v
it
, (v, v
i
0
), v
i
0
is a directed circuit.
Hence, d
(v
i
0
) = 0. Using a similar technique, we can show that d
+
(v
i
k
) = 0 as well.
If G = (V, E) is a digraph with n vertices, then a labeling of the vertices with an injective
function : V 1, . . . , n which satises the condition (u) < (v) whenever (u, v) is an
arc in G is known as topological sorting.
Theorem 3.5. We can sort the vertices of a digraph topologically if and only if the graph is
acyclic.
Proof. If the digraph is cyclic, then obviously we can not sort the vertices topologically.
If the digraph G is acyclic, then we can sort the vertices in the following manner:
2
1. We choose a vertex v which is a sink. It exists by Theorem 3.4. We set (v) n,
G Gv and n n 1.
2. If there is just one vertex v in G, set (v) 1. Otherwise, go back to step #1.
2
This is known as Marimonts Algorithm. The algorithm itself contains other items, too. The original refer-
ence is MARIMONT, R.B.: A New Method of Checking the Consistency of Precedence Matrices. Journal of the
Association for Computing Machinery 6 (1959), 164171.
Chapter 4
Matrices and Vector Spaces of Graphs
4.1 Matrix Representation of Graphs
The adjacency matrix of the graph G = (V, E) is an n n matrix D = (d
ij
), where n is the
number of vertices in G, V = v
1
, . . . , v
n
and
d
ij
= number of edges between v
i
and v
j
.
In particular, d
ij
= 0 if (v
i
, v
j
) is not an edge in G. The matrix D is symmetric, i.e. D
T
= D.
Example.
D =
_
_
_
_
_
_
0 2 1 0 0
2 1 0 1 0
1 0 3 0 0
0 1 0 0 0
0 0 0 0 0
_
_
_
_
_
_
v
1
v
3
v
5
v
4
v
2
Obviously, an adjacency matrix denes a graph completely up to an isomorphism.
The adjacency matrix of a directed graph G is D = (d
ij
), where
d
ij
= number of arcs that come out of vertex v
i
and go into vertex v
j
.
Example.
D =
_
_
_
_
0 1 0 0
1 0 0 0
0 0 0 0
2 1 0 1
_
_
_
_
v
1
v
2
v
4
v
3
The all-vertex incidence matrix of a non-empty and loopless graph G = (V, E) is an n m
matrix A = (a
ij
), where n is the number of vertices in G, m is the number of edges in G and
a
ij
=
_
1 if v
i
is an end vertex of e
j
0 otherwise.
34
CHAPTER 4. MATRICES AND VECTOR SPACES OF GRAPHS 35
Example.
e
1
e
2
e
3
e
4
A =
_
_
_
_
_
_
1 1 1 0
1 1 0 1
0 0 1 0
0 0 0 1
0 0 0 0
_
_
_
_
_
_
v
1
v
2
v
3
v
4
v
5
v
1
v
2
v
3
v
4
v
5
e
3
e
4
e
2
e
1
The all-vertex incidence matrix of a non-empty and loopless directed graph G is A = (a
ij
),
where
a
ij
=
_
_
1 if v
i
is the initial vertex of e
j
1 if v
i
is the terminal vertex of e
j
0 otherwise.
Example.
e
1
e
2
e
3
e
4
e
5
A =
_
_
_
_
1 1 1 1 0
1 1 0 0 1
0 0 0 0 0
0 0 1 1 1
_
_
_
_
v
1
v
2
v
3
v
4
v
1
v
2
v
4
v
3
e
3
e
4
e
1
e
2
e
5
Since every column of an all-vertex incidence matrix contains exactly two non-zero num-
bers, two ones, we can remove a row and still have enough information to dene the graph. The
incidence matrix of a graph is obtained by removing a row from the all-vertex incidence matrix.
It is not unique because there are n possible rows to remove. The vertex corresponding to the
row removed is called the reference vertex.
Similarly, every column in the all-vertex incidence matrix of a digraph contains exactly two
non-zero numbers, +1 and 1. We can remove a row from the all-vertex incidence matrix and
obtain the incidence matrix. Notice that the rows of an all-vertex incidence matrix are linearly
dependent because the sum of rows is a zero vector.
Theorem 4.1. The determinant of an incidence matrix of a nontrivial tree is 1, regardless of
whether the tree is a directed graph or not.
Proof. We use induction on n, the number of vertices in the tree.
Induction Basis: n = 2 and it is obvious.
Induction Hypothesis: The theorem is true for n k. (k 2)
Induction Statement: The theorem is true for n = k + 1.
Induction Statement Proof: Let T be a tree which has k + 1 vertices and let A be an (arbi-
trary) incidence matrix of T. T has at least two pendant vertices (Theorem 2.3). We choose a
pendant vertex v
i
which is not the reference vertex of Aand the edge e
t
which is incident on v
i
.
Then,
a
it
= ()1 and a
ij
= 0, when j ,= t.
We expand the determinant of [A[ by the i
th
row:
[A[ = ()(1)
i+t
[A
[,
where A
is the minor corresponding to a

it
. We write T
= T v
i
which is also a tree (v
i
is a
pendant vertex). We use the induction hypothesis to get [A
[ = 1 because A
is obviously an
incidence matrix of T
.
Corollary. If the digraph G has no loops, then the rank of its all-vertex incidence matrix is
(G).
Proof. If we rearrange the rows or columns of the all-vertex incidence matrix, the rank of the
matrix will not change. Let us rearrange the vertices and arcs to group them by components.
Then, the all-vertex incidence matrix is a block diagonal matrix in which each block is an all-
vertex incidence matrix of a component.
_
_
_
_
_
_
_
_
_
_
_
1
st
compo-
nent
2
nd
compo-
nent
O
O
.
.
.
k
th
compo-
nent
_
_
_
_
_
_
_
_
_
_
_
We denote n
i
as the number of vertices in the i
th
component. Every component has a spanning
tree whose incidence matrix has a nonzero determinant by Theorem 4.1, i.e. the matrix is not
singular. The all-vertex incidence matrix of the i
th
component is obtained by adding columns
and one row to an incidence matrix of the corresponding spanning tree. The row added is
linearly dependent of other rows so that the rank of this matrix is the same as the rank of the
incidence matrix (= n
i
1). Notice that in the special case when a component is trivial, the
rank is zero = 1 1. Therefore,
rank of A = sum of the ranks of the components
= (n
1
1) + + (n
k
1)
= n
1
+ +n
k
. .
= n
k = (G).
Remark. From this proof, we can also get a basis for the row space and the column space of
the all-vertex incidence matrix. The columns corresponding to the branches of the spanning
forest of G are a basis of the column space. We can get a basis of the row space by removing
one row out of each component block.
4.2 Cut Matrix
If all the cuts of a nontrivial and loopless graph G = (V, E) are I
1
, . . . , I
t
, then the cut matrix
of G is a t m matrix Q = (q
ij
), where m is the number of edges in G and
q
ij
=
_
1 if e
j
I
i
(the cut is interpreted as an edge set)
0 otherwise.
Example. For the graph
v
2
v
3
e
2
e
3
e
4
e
1
v
1
the cuts are I
1
= e
1
, e
4
, I
2
= e
2
, e
3
, e
4
and I
3
= e
1
, e
2
, e
3
. The cut matrix is
e
1
e
2
e
3
e
4
Q =
_
_
1 0 0 1
0 1 1 1
1 1 1 0
_
_
I
1
I
2
I
3
Remark. If the graph has n vertices, then it has
1
2
(2
n
2) = 2
n1
1 cuts. Usually, there are
not this many distinct edge sets. For the cut matrix, we only take one cut corresponding to an
edge set so that there would not be repeated rows. Even so, there are usually too many rows.
If G is a nontrivial and loopless digraph, then we assign an arbitrary direction to every cut
V
1
, V
2
: the orientation of V
1
, V
2
is from V
1
to V
2
. In other words, we consider oriented cuts
and we pick only one direction from the two possibilities. Then, the cut matrix Q = (q
ij
) is
q
ij
=
_
_
1 if e
j
I
i
and they are in the same direction
1 if e
j
I
i
and they are in opposite directions
0 otherwise.
Example. For the digraph
v
1
v
2
v
4
v
3
e
3
e
4
e
1
e
2
e
5
the different cuts (interpreted as edge sets) are I
1
= e
1
, e
2
, e
3
, e
4
(in the direction of e
1
),
I
2
= e
3
, e
4
, e
5
3
), I
3
= e
1
, e
2
, e
5
1
) and I
4
= .
The cut matrix is
e
1
e
2
e
3
e
4
e
5
Q =
_
_
_
_
1 1 1 1 0
0 0 1 1 1
1 1 0 0 1
0 0 0 0 0
_
_
_
_
I
1
I
2
I
3
I
4
Since v, V v is a cut for every vertex v, rows of the all-vertex incidence matrix are
rows of Q. If we are dealing with directed graphs, then these rows may have to be multiplied
by 1.
Theorem 4.2. Every row of the cut matrix of a digraph can be expressed in two different ways
as a linear combination of the rows of the all-vertex incidence matrix. The non-zero coefcients
are either all = +1 or all = 1.
Proof. Let Q be the cut matrix of a digraph G = (V, E) and let A be the all-vertex incidence
matrix. Let V
1
, V
2
(note that it is oriented) be the cut corresponding to the i
th
row of Q.
Reindexing if needed, we can assume that
V
1
= v
1
, . . . , v
r
and V
2
= v
r+1
, . . . , v
n
.
We write
q
i
= i
th
row of Q and a
t
= t
th
row of A.
We show that
q
i
=
r
t=1
a
t
=
n
t=r+1
a
t
,
which proves the theorem. Let (v
p
, v
q
) = e
k
be the k
th
arc of G. Then,
a
pk
= k
th
element of the vector a
p
= 1,
a
qk
= k
th
element of the vector a
q
= 1
and
a
jk
= 0 if j ,= p, q.
We get four cases:
v
p
V
1
and v
q
V
2
: Now p r and q r + 1 so q
ik
= 1 and
q
ik
=
r
t=1
a
tk
=
n
t=r+1
a
tk
.
v
p
V
2
and v
q
V
1
: Now p r + 1 and q r so q
ik
= 1 and
q
ik
=
r
t=1
a
tk
=
n
t=r+1
a
tk
.
v
p
V
1
and v
q
V
1
: Now p r and q r so q
ik
= 0 and
q
ik
=
r
t=1
a
tk
= a
r+1,k
. .
=0
a
nk
..
=0
.
v
p
V
2
and v
q
V
2
: Now p r + 1 and q r + 1 so q
ik
= 0 and
q
ik
= a
1k
..
=0
+ + a
rk
..
=0
=
n
t=r+1
a
tk
.
The statements above are valid for every k.
Example. (Continuing from the previous example) The corresponding row of I
1
is
(1, 1, 1, 1, 0) = (1, 1, 1, 1, 0) = (1, 1, 0, 0, 1) (0, 0, 0, 0, 0) (0, 0, 1, 1, 1).
Corollary. The rank of the cut matrix of a digraph G is (G).
Proof. The all-vertex incidence matrix A of G is also a submatrix of the cut matrix Q of G.
Then, (by Corollary of Theorem 4.1)
rank(Q) rank(A) = (G).
On the other hand, by Theorem 4.2, every row of Q can be expressed as a linear combination
of the rows of A. Therefore,
rank(Q) = rank(A) = (G).
Another consequence is that the cut matrix Q can be expressed as
Q = A
1
A,
where the elements of A
1
are 0 or 1. In addition, the matrix A
1
can be constructed from the
process in the proof of Theorem 4.2.
If the graph G is connected, then it has a spanning tree T and an associated fundamental cut
set. The fundamental cut sets are also cuts (when cuts are interpreted as edge sets). Therefore,
the cut matrix Q of G has a submatrix Q
f
that corresponds to these fundamental cut sets.
This matrix is called the fundamental cut set matrix. Similarly, the connected digraph G has
a fundamental cut set matrix: if we interpret a fundamental cut set as a set, then the direction
of the cut is chosen to be the same as the direction of the corresponding branch of T. If we
rearrange the edges of G so that we have the branches rst and sort the fundamental cut sets in
the same order, then we get the fundamental cut set matrix in the form
Q
f
=
_
I
n1
Q
fc
_
,
where I
n1
is the identity matrix with n 1 rows. The rank of Q
f
is thus n 1 = (G).
Example. (Continuing from the previous example) We left out vertex v
3
so we get a connected
digraph. We choose the spanning tree
v
1
v
2
v
4
e
3
e
1
T:
The fundamental cut sets are I
2
= e
3
, e
4
, e
5
3
) and I
3
= e
1
, e
2
, e
5
(in
the direction of e
1
). Then,
e
1
e
3
e
2
e
4
e
5
Q
f
=
_
1 0 1 0 1
0 1 0 1 1
_
I
3
I
2
and
e
1
e
2
e
3
e
4
e
5
Q =
_
_
1 1 1 1 0
0 0 1 1 1
1 1 0 0 1
_
_
4.3 Circuit Matrix

We consider a loopless graph G = (V, E) which contains circuits. We enumerate the circuits of
G: C
1
, . . . , C
. The circuit matrix of G is an m matrix B = (b

ij
) where
b
ij
=
_
1 if the arc e
j
is in the circuit C
i
0 otherwise
(as usual, E = e
1
, . . . , e
m
).
The circuits in the digraph G are oriented, i.e. every circuit is given an arbitrary direction
for the sake of dening the circuit matrix. After choosing the orientations, the circuit matrix of
G is B = (b
ij
) where
b
ij
=
_
_
1 if the arc e
j
is in the circuit C
i
and they in the same direction
1 if the arc e
j
is in the circuit C
i
and they are in the opposite direction
0 otherwise.
Example. For the directed graph
v
1
e
4
e
1
v
2
e
2
e
3
v
3
the circuits are
v
1
e
4
e
1
v
2
e
3
v
3
C
1
v
1
e
4
e
1
v
2
e
2
v
3
C
2
v
2
e
2
e
3
v
3
C
1
and the circuit matrix is
e
1
e
2
e
3
e
4
B =
_
_
1 0 1 1
1 1 0 1
0 1 1 0
_
_
C
1
C
2
C
3
If the graph G is connected and contains at least one circuit, then it has a cospanning tree
T
and the corresponding fundamental circuits. By choosing the corresponding rows of the
circuit matrix B, we get an (mn + 1) m matrix B
f
, called the fundamental circuit matrix.
Similarly, a connected digraph G with at least one circuit has a fundamental circuit matrix: the
direction of a fundamental circuit is the same as the direction of the corresponding link in T
.
When we rearrange the edges of Gso that the links of T
come last and sort the fundamental

circuits in the same order, the fundamental circuit matrix takes the form
B
f
=
_
B
ft
I
mn+1
_
,
where I
mn+1
is the identity matrix with mn+1 rows. The rank of B
f
is thus mn+1 = (G)
and the rank of B is mn + 1.
Example. (Continuing from the previous example) We left out vertex v
3
so we get a connected
digraph (see p.34) and we chose the spanning tree
v
1
v
2
v
4
e
3
e
1
T:
The fundamental circuits are
v
1
v
2
e
1
e
2
C
1
v
1
v
4
e
3
e
4
C
2
v
1
v
2
v
4
e
3
e
1
e
5
C
3
and
e
1
e
3
e
2
e
4
e
5
B
f
=
_
_
1 0 1 0 0
0 1 0 1 0
1 1 0 0 1
_
_
C
1
C
2
C
3
Theorem 4.3. An oriented cut and an oriented circuit of a digraph have an even number of
common arcs. Half of these arcs have the same direction in the cut and in the circuit, and the
remaining arcs have opposite directions in the cut and in the circuit.
Proof. Compare to the proof of Theorem 2.6.
Theorem 4.4. For a digraph, BQ
T
= O (zero matrix).
Proof. By the previous theorem, half of the nonzero numbers in the dot product corresponding
to each element of BQ
T
are +1. The remaining nonzero numbers are 1. Therefore, the dot
product is = 0.
Theorem 4.5. If the digraph G contains at least one circuit, then the rank of its circuit matrix
B is (G). Furthermore, if G is connected, then the circuit matrix B can be expressed as
B = B
2
B
f
, where the matrix B
2
consists of 0s and 1s, and the cut matrix Qcan be expressed
as Q = Q
1
Q
f
, where the matrix Q
1
consists of 0s and 1s.
Proof. First we consider the case when G is connected. We choose a spanning tree T of G and
rearrange the m edges of G so that the branches of T come rst and the links of T
come last.
We sort the fundamental cut sets in the same order as the branches and links. Then,
Q
f
=
_
I
n1
Q
fc
_
and B
f
=
_
B
ft
I
mn+1
_
.
The blocks of B can be constructed in a similar way:
B =
_
B
1
B
2
_
.
Since Q
f
is a submatrix of Q and B
f
is a submatrix of B, it follows from Theorem 4.4 that
O = B
f
Q
T
f
=
_
B
ft
I
mn+1
_ _
I
n1
Q
fc
_
T
=
_
B
ft
I
mn+1
_
_
I
n1
Q
T
fc
_
= B
ft
I
n1
+I
mn+1
Q
T
fc
= B
ft
+Q
T
fc
.
Hence
B
ft
= Q
T
fc
.
Furthermore, since Q
f
is a submatrix of Q, we can use the same theorem to get
O = BQ
T
f
=
_
B
1
B
2
_ _
I
n1
Q
fc
_
T
=
_
B
1
B
2
_
_
I
n1
Q
T
fc
_
= B
1
I
n1
+B
2
Q
T
fc
= B
1
B
2
B
ft
.
Hence
B =
_
B
2
B
ft
B
2
_
= B
2
_
B
ft
I
mn+1
_
= B
2
B
f
,
as claimed. In the same way, Q can be expressed as Q = Q
1
Q
f
, as claimed, which is clear
anyway since the rank of Q is n 1 and its elements are 0s and 1s.
Every rowof Bis a linear combination of the rows corresponding to the fundamental circuits
and the rank of B is at most equal to the rank of B
f
= m n + 1. On the other hand, as we
pointed out earlier, the rank of B is mn + 1. Thus, rank(B) = mn + 1 (= (G)) for a
connected digraph.
In the case of a disconnected digraph G(which contains at least one circuit), it is divided into
components (k 2 components) and the circuit matrix B is divided into blocks corresponding
to the components (compare to the proof of the corollary of Theorem 4.1), in which case
rank(B) =
k
i=1
(m
i
n
i
+ 1) = mn +k = (G).
Notice that the proof also gives the formula, B
ft
= Q
T
fc
, which connects the fundamental
cut matrix and the fundamental circuit matrix.
4.4 An Application: Stationary Linear Networks
A stationary linear network is a directed graph G that satises the following conditions:
1. G is connected.
2. Every arc of G belongs to some circuit and there are no loops in G.
3. Every arc e
j
in G is associated with a number i
j
called the through-quantity or ow. If
there are m arcs in G, then we write
i =
_
_
_
i
1
.
.
.
i
m
_
_
_
(through-vector).
4. Every vertex v
i
in G is associated with a number p
i
called the potential. Furthermore, the
across-quantity or potential difference of the arc e
j
= (v
i
1
, v
i
2
) is
u
j
= p
i
2
p
i
1
.
If there are n vertices and m arcs in G, then we write
p =
_
_
_
p
1
.
.
.
p
n
_
_
_
and u =
_
_
_
u
1
.
.
.
u
m
_
_
_
(potential vector and across-vector). (Potentials are rarely needed.)
5. Every arc e
j
is one of the following:
(a) a component
1
, for which there is an associated number r
j
. r
j
is constant (,= 0)
(stationarity) and the following equation links the quantities:
u
j
= i
j
r
j
(linearity).
(b) a through-source, for which the through-quantity i
j
is xed.
(c) an across-source, for which the across-quantity u
j
is xed.
6. (Kirchhoff s Through-Quantity Law) The sum of the through-quantities of an oriented cut
of G is zero when the cut is interpreted as an edge set and the sign of a through-quantity
is changed if the directions of a cut and an arc are different.
7. (Kirchhoff s Across-Quantity Law) The sum of the across-quantities of an oriented circuit
of G is zero when the sign of an across-quantity is changed if the directions of a circuit
and an arc are different.
1
Not to be confused with a component of a graph!
Example. A typical stationary linear network is an electrical circuit with linear resistors, con-
stant current sources and constant voltage sources. The components are resistors and r
j
are the
resistances. Equation 5.(a) is Ohms Law.
We take a spanning tree T of a stationary linear network G, its fundamental cut matrix Q
f
and its fundamental circuit matrix B
f
. Let us rearrange the arcs in these matrices and vectors i
and u like we did before. That is, the branches of T will come rst followed by the links of T
.
Kirchhoffs Laws can then be written as
Qi = 0 and Bu = 0.
On the other hand, the rows of the fundamental cut matrix Q
f
span all the rows of Q, and
similarly rows of the fundamental circuit matrix B
f
span the rows of B. Then, Kirchhoffs
Laws can also be written as
Q
f
i = 0
n1
and B
f
u = 0
mn+1
.
Let us form the diagonal matrices K = k
1
, . . . , k
m
and L =
1
, . . . ,
m
, where
k
j
=
_
_
r
j
if e
j
is a component
1 if e
j
is a through-source
0 if e
j
is an across-source
and
j
=
_
_
1 if e
j
is a component
0 if e
j
is a through-source
1 if e
j
is an across-source,
and the m-vector s = (s
1
, . . . , s
m
)
T
, where
s
j
=
_
_
0 if e
j
is a component
i
j
if e
j
is a through-source
u
j
if e
j
is an across-source.
Then, all the information can be expressed as a system of linear equations
_
_
K L
Q
f
O
(n1)m
O
(mn+1)m
B
f
_
_
_
i
u
_
=
_
_
s
0
n1
0
mn+1
_
_
,
known as the fundamental equations. The through and across quantities can be solved (ideally)
if r
j
and the sources are given.
Remark. The same procedure can be applied to form state (differential) equations for dynamic
networks, which have nonstationary components.
The matrix of this system of linear equations does not have to be nonsingular and the system
does not even have to have a unique solution at all. For example, in the matrix above, we can
easily see that it is singular if some circuit only consists of across-sources or if some cut only
consists of through-sources. As a matter of fact, this is the only case when the through and
across quantities are not dened uniquely if the constants r
j
are real numbers with the same
sign (and often otherwise too).
We choose a specic spanning tree T to explore these concepts more carefully:
Lemma. If no cut of G consists of only through-sources and no circuit of G consists of only
across-sources, then G has a spanning tree T such that every across-source is a branch of T
and every through-source is a link of T
.
Proof. If G satises the hypothesis, then we rst choose a digraph M which has every vertex
and across-source (arc) of G. There are no circuits in this digraph. Then we add components to
M one by one and try to come up with a spanning tree. If this fails at some point, then G has a
cut with only through-sources, which is impossible.
Now let us assume that no cut of G consists of only through-sources and no circuit of G
consists of only across-sources. We use the spanning tree T mentioned in the lemma. We
rearrange the arcs of G so that (as before) the branches of T come rst. Within these branches,
the across-sources come rst followed by components. Similarly, the links are rearranged so
that the components come rst and the through-sources come last.
The system of 2m equations can then be written as
_
_
_
_
_
_
_
_
_
_
_
_
O O O O I O O O
O R
1
O O O I O O
O O R
2
O O O I O
O O O I O O O O
I O Q
11
Q
12
O O O O
O I Q
21
Q
22
O O O O
O O O O B
11
B
12
I O
O O O O B
21
B
22
O I
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
i
1
i
2
i
3
i
4
u
1
u
2
u
3
u
4
_
_
_
_
_
_
_
_
_
_
_
_
=
_
_
_
_
_
_
_
_
_
_
_
_
s
1
0
0
s
2
0
0
0
0
_
_
_
_
_
_
_
_
_
_
_
_
across-sources (in branches)
components (in branches)
components (in links)
through-sources (in links)
fundamental cut sets (across-sources)
fundamental cut sets (components)
fundamental circuits (components)
fundamental circuits (through-sources)
where the Is are identity matrices of the right dimensions, the Os are zero matrices of the right
dimensions and the 0s are zero vectors of the right dimensions.
Remark. Here we assume that G has all these four types of arcs (across-source branch, com-
ponent branch, through-source link and component link). In other cases (for example, when
there are no through-sources), we leave the corresponding rows, columns and elements out of
the system of equations. Other cases are treated in a similar way.
Solving the equations, we get
u
1
= s
1
, u
2
= R
1
i
2
, u
3
= R
2
i
3
, i
4
= s
2
which leaves this system of equations:
_
_
_
_
I O Q
11
O
O I Q
21
O
O B
12
R
1
R
2
O
O B
22
R
1
O I
_
_
_
_
_
_
_
_
i
1
i
2
i
3
u
4
_
_
_
_
=
_
_
_
_
Q
12
s
2
Q
22
s
2
B
11
s
1
B
21
s
1
_
_
_
_
.
Thus,
i
1
= Q
11
i
3
Q
12
s
2
, i
2
= Q
21
i
3
Q
22
s
2
, u
4
= B
22
R
1
i
2
B
21
s
1
.
From the results in the previous section, we get
B
ft
=
_
B
11
B
12
B
21
B
22
_
= Q
T
fc
=
_
Q
11
Q
12
Q
21
Q
22
_
T
=
_
Q
T
11
Q
T
21
Q
T
12
Q
T
22
_
.
Therefore, B
11
= Q
T
11
and B
12
= Q
T
21
. Finally, there is only one set of equations for i
3
:
(Q
T
21
R
1
Q
21
+R
2
)i
3
= Q
T
11
s
1
Q
T
21
R
1
Q
22
s
2
.
The matrix
2
of this system of equations can written as
Q
T
21
R
1
Q
21
+R
2
=
_
Q
T
21
I
_
_
R
1
O
O R
2
__
Q
21
I
_
.
We can see that it is not singular if the diagonal elements of R
1
and R
2
are all positive or all
negative.
Therefore, we get
Theorem 4.6. If the constants r
j
are real numbers with the same sign, then the fundamental
equations of the stationary linear network have a unique solution exactly when no cut of the
network consists of only through-sources and no circuit of the network consists of only across-
sources.
From the theorem above, we notice that the number of equations we have to solve (numerically)
is considerably fewer than 2m.
Example. A mono-frequency AC circuit with passive elements (resistors, capacitors and induc-
tors) can also be modelled as a stationary linear network (Theorem 4.6 does not apply). In the
circuit below, the component values are R = 10 , C = 100 F, L = 10 mH and the current
source is
I = 10 cos(1000t) A.
I
R
R
L
C
The complex current of the source is 10e
j1000t
, where j is the imaginary unit. The (angular)
frequency is = 1000 rad/s. There are no voltage sources. The corresponding digraph is
2
This matrix is called the admittance matrix. Similarly, the impedance matrix can be constructed from the
blocks of the fundamental circuit matrix
v
3
e
5
v
1
e
3
v
2
e
2
e
4
e
1
The voltages and currents written as complex exponentials are i
k
= I
k
e
j1000t
and u
k
= U
k
e
j1000t
.
In particular, the current source is i
5
= s
5
= 10e
j1000t
. We get r
k
from the familiar formulae
from electrical circuit analysis:
r
1
= r
4
= R = 10 , r
3
=
1
jC
= 10j , r
2
= jL = 10j.
We choose the arcs e
1
and e
2
as the branches of the spanning tree T. Because of the linearity
of the system of equations, the exponential factors e
j1000t
cancel out and we get
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
10 0 0 0 0 1 0 0 0 0
0 10j 0 0 0 0 1 0 0 0
0 0 10j 0 0 0 0 1 0 0
0 0 0 10 0 0 0 0 1 0
0 0 0 0 1 0 0 0 0 0
1 0 1 1 1 0 0 0 0 0
0 1 0 1 0 0 0 0 0 0
0 0 0 0 0 1 0 1 0 0
0 0 0 0 0 1 1 0 1 0
0 0 0 0 0 1 0 0 0 1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
I
1
I
2
I
3
I
4
I
5
U
1
U
2
U
3
U
4
U
5
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0
0
0
0
10
0
0
0
0
0
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
.
Notice that we have left out the across-sources because there are none. This system is easily
solved using computer programs, e.g. MATLAB:
H=[-10 0 0 0 0 1 0 0 0 0 ;
0 -10
*
j 0 0 0 0 1 0 0 0 ;
0 0 10
*
j 0 0 0 0 1 0 0 ;
0 0 0 -10 0 0 0 0 1 0 ;
0 0 0 0 1 0 0 0 0 0 ;
1 0 1 1 1 0 0 0 0 0 ;
0 1 0 -1 0 0 0 0 0 0 ;
0 0 0 0 0 -1 0 1 0 0 ;
0 0 0 0 0 -1 1 0 1 0 ;
0 0 0 0 0 -1 0 0 0 1];
s=[0 0 0 0 10 0 0 0 0 0];
UV=inv(H)
*
s;
[UV angle(UV) abs(UV)]
ans =
-6.0000 + 2.0000i 2.8198 6.3246
-2.0000 + 4.0000i 2.0344 4.4721
-2.0000 - 6.0000i -1.8925 6.3246
-2.0000 + 4.0000i 2.0344 4.4721
10.0000 0 10.0000
-60.0000 +20.0000i 2.8198 63.2456
-40.0000 -20.0000i -2.6779 44.7214
-60.0000 +20.0000i 2.8198 63.2456
-20.0000 +40.0000i 2.0344 44.7214
-60.0000 +20.0000i 2.8198 63.2456
Thus, for example, the complex voltage across the current source is
u
5
= U
5
e
j1000t
= 63.25e
j(1000t+2.82)
and the real voltage is 63.25 cos(1000t + 2.82) V.
Kirchhoffs Through-Quantity Law can also be written in the form
Ai = 0
n
,
where A is the all-vertex incidence matrix of G. Furthermore,
A
T
p = u.
Hence
u i = u
T
i = p
T
Ai = 0.
This result only depends on the structure of the digraph G (through the all-vertex incidence
matrix). Now we get the famous theorem:
Theorem 4.7. (Tellegen) If two stationary linear networks have the same digraph with corre-
sponding through-vectors i
1
and i
2
as well as corresponding across-vectors u
1
and u
2
, then
u
1
i
2
= 0 and u
2
i
1
= 0.
If we apply this to the case when the two networks are exactly the same (= G), then we get
u i = 0,
known as the Law of Conservation of Energy.
Remark. More details on this subject can be found e.g. in SWAMY & THULASIRAMAN or
VG, as well as DOLAN & ALDOUS.
4.5 Matrices over GF(2) and Vector Spaces of Graphs
The set 0, 1 is called a eld (i.e. it follows the same arithmetic rules as real numbers) if
addition and multiplication are dened as follows:
+ 0 1
0 0 1
1 1 0
0 1
0 0 0
1 0 1
In this case 1 = 1 and 1
1
= 1. This is the eld GF(2).
If we think of the elements 0 and 1 of the all-vertex incidence, cut, fundamental cut, circuit
and fundamental circuit matrices of a (undirected) graph as elements of the eld GF(2), then
Theorems 4.1, 4.2, 4.4, 4.5 and their corollaries also apply to undirected graphs. (Keep in
mind that 1 = 1 in the eld GF(2).) The proofs are the same.
For undirected graphs, the vector spaces are over the eld GF(2). For directed graphs, the
vector spaces are real (i.e. over the eld R). The row space of the cut matrix of a (di)graph is
the cut space. Similarly, the row space of the circuit matrix is the circuit space. The dimension
of the cut space is the rank of the (di)graph and the dimension of the circuit space is the nullity
of the (di)graph. Furthermore, the cut space and the circuit space are orthogonal complements.
(All of these statements follow directly from the results above.)
Often, we deal with the above mentioned spaces through subgraphs, i.e. we identify a vector
with the subgraph generated by the corresponding arcs. In the case of undirected graphs, the
addition of GF(2) vectors corresponds to the ring sum operation.
Chapter 5
Graph Algorithms
5.1 Computational Complexity of Algorithms
The complexity of a problem is related to the resources required to compute a solution as a
function of the size of the problem. The size of a problemis measured by the size of the input N,
and the resources required are usually measured by time (number of steps) and space (maximum
amount of memory measured appropriately). Decision problems or yes-or-no questions are very
common. Read HOPCROFT & ULLMAN for classical complexity theory.
To make computational complexities comparable, we need to agree on some specic math-
ematical models for algorithms. For example, consider computing with Turing Machines and
refer to courses in Formal Languages, Theory of Automata and Mathematical Logic. We have
deterministic and nondeterministic version of algorithm models. In the deterministic version,
there are no choices to be made. In the nondeterministic version, there is a choice to be made
somewhere on the way. For a nondeterministic algorithm, we have to make the following as-
sumptions so that we can actually solve problems:
1. The algorithm terminates at some point no matter how we choose the steps.
2. The algorithm can terminate without yielding a solution.
3. When the algorithm terminates and yields a solution, the solution is correct (it is possible
to have more than one solution).
4. For decision problems, if the algorithm fails to give a positive answer (yes), then the
answer is interpreted to be negative (no).
5. If the problem is to compute a value, then the nondeterministic algorithm has to give a
solution for every input (value of the function).
Nondeterministic algorithms are best treated as verication procedures for problems rather than
procedures for producing answers.
Computational complexity is considered asymptotically, that is for large problems, time
or space complexities that differ by constant coefcients are not distinguished because linear
acceleration and compression of space are easy to perform in any kind of algorithm model.
Although the choice of an algorithm model has a clear impact on the complexity, it is not an
essential characteristic, i.e. it does not change the complexity class. Often, we use the big-O
notation for complexities. O(f(N)) refers to the class of functions g(N) such that if N N
0
holds, then [g(N)[ Cf(N) holds, where C is a constant.
50
CHAPTER 5. GRAPH ALGORITHMS 51
Without exploring algorithm models any further, we dene a couple of important complex-
ity classes. The time complexity class T (deterministic polynomial time problems) consists of
problems of (input) size N where it takes at most p(N) steps to solve the problem using deter-
ministic algorithms. p(N) is some problem dependent polynomial of N. The time complexity
class AT (nondeterministic polynomial time problems) consists of problems of size N where it
takes at most p(N) steps to solve the problem using nondeterministic algorithms. Once again,
p(N) is some problem dependent polynomial of N.
Time complexity class coAT (complements of nondeterministic polynomial time prob-
lems) consists of decision problems whose complements are in AT. (The complement of a
problem is obtained by swapping the positive and the negative answer.)
Obviously, T AT and (for decision problems) T coAT. Whether or not the inclu-
sion is proper is an open problem, actually quite a famous problem. It is widely believed that
both of the inclusions are proper. It is not known if the following holds for decision problems:
AT = coAT or T = AT coAT
Most researchers believe that they do not hold.
The space complexity class ToT/(c (deterministic polynomial space problems) consists
of problems of (input) size N where it takes at most p(N) memory units to solve the prob-
lem using deterministic algorithms. p(N) is some problem dependent polynomial of N. The
space complexity class AToT/(c (nondeterministic polynomial space problems) consists of
problems of size N where it takes at most p(N) memory units to solve the problem using non-
deterministic algorithms. Once again, p(N) is some problem dependent polynomial of N. It is
known that
AT ToT/(c = AToT/(c,
but it is not known whether the inclusion is proper or not.
An algorithm may include some ideally generated random numbers. The algorithm is then
called probabilistic or stochastic. The corresponding polynomial time complexity class is BTT
(random polynomial time problems or bounded-error probabilistic polynomial time problems).
Some stochastic algorithms may fail occasionally, that is, they produce no results and terminate
prematurely. These algorithms are called Las Vegas algorithms. Some stochastic algorithms
may also produce wrong answers (ideally with a small probability). These kind of algorithms
are called Monte Carlo algorithms. Some stochastic algorithms seldom yield exact solutions.
Nevertheless, they give accurate approximate solutions with high probability. These kind of
algorithms are called approximation algorithms.
The task of an algorithmmay be to convert a problemto another. This is known as reduction.
If problem A can be reduced to another problem B by using a (deterministic) polynomial time
algorithm, then we can get a polynomial time algorithm for problem A from a polynomial time
algorithm for B. A problem is AT-hard if every problem in AT can be reduced to it by a
polynomial time algorithm. AT-hard problems are AT-complete if they are actually in AT.
AT-complete problems are the worst kind. If any problem in AT could be shown to be
deterministic polynomial time, then every problem in AT would be in T and T = AT. Over
one thousand AT-complete problems are known currently.
The old division of problems into tractable and intractable means that T problems are
tractable and others are not. Because we believe that T , = AT in general, AT-complete
problems are intractable. In the following, graph algorithms are either in T or they are ap-
proximations of some more demanding problems. The size of an input can be for example the
number of nonzero elements in an incidence matrix, the number of vertices n or the number of
edges m or some combination of n and m.
5.2 Reachability: Warshalls Algorithm
We only deal with directed graphs in this section. The results also hold for undirected graphs
if we interpret an edge as a pair of arcs in opposite directions.
Problem. We are given an adjacency matrix of the digraph G = (V, E). We are to construct
the reachability matrix R = (r
ij
) of G, where
r
ij
=
1 if G has a directed v
i
v
j
path
0 otherwise.
(Note that V = v
1
, . . . , v
n
.) In particular, we should note that if r
ii
= 1, then v
i
is in a
directed circuit.
Warshalls Algorithm constructs a series of n n matrices E
1
, . . . , E
n
where
1. elements of E
i
are either zero or one.
2. E
i
E
i+1
(i = 0, . . . , n 1) (comparison is done element by element).
3. E
0
is obtained from the adjacency matrix Dby replacing the positive elements with ones.
4. E
n
= R.
The algorithm is presented as a pseudocode:
procedure Warshall
begin
E := E
0
for i := 1 to n do
for j := 1 to n do
if (E)
ji
= 1 then for k := 1 to n do
(E)
jk
:= max((E)
jk
, (E)
ik
)
od
od
end
In this case, the maximizing operation is sometimes called the Boolean sum:
max 0 1
0 0 1
1 1 1
Let us show that Warshalls Algorithm gives us the desired results. Let E
i
denote the value
of E after i steps.
Statement. (i) If there is a directed path from v
s
to v
t
such that apart from v
s
and v
t
, the path
only includes vertices in the set v
1
, . . . , v
i
, then (E
i
)
st
= 1.
(ii) If vertex v
s
belongs to a directed circuit whose other vertices are in the set v
1
, . . . , v
i
,
then (E
i
)
ss
= 1.
Proof. We will use induction on i.
Induction Basis: i = 1. (E
1
)
st
= 1 if (E
0
)
st
= 1, or (E
0
)
s1
= 1 and (E
0
)
1t
= 1. We have
one of the following cases:
s t: v
s
v
t
v
s
v
1
v
t
s = t: v
s
v
s
v
1
Induction Hypothesis: The statement is true for i < . ( 2)
Induction Statement: The statement is true for i = .
Induction Statement Proof: Let us handle both statements together. The proof for (ii) is
given in square brackets. We have two cases:
v
belongs to the directed path [resp. directed circuit] but ,= s, t [resp. ,= s]. Then, we
use the Induction Hypothesis:
(E
1
)
s
= 1 and (E
1
)
t
= 1 [resp. (E
1
)
s
= 1 and (E
1
)
s
= 1],
so (E
)
st
= 1 [resp. (E
)
ss
= 1].
v
is either v
s
or v
t
[resp. v
is v
s
] or it does not belong to the directed path [resp. directed
circuit] at all. Then, by the Induction Hypothesis
(E
1
)
st
= 1 [resp. (E
1
)
ss
= 1],
so (E
)
st
= 1 [resp. (E
)
ss
= 1].
In Warshalls Algorithm, the maximizing operation is performed at most n
3
times.
5.3 Depth-First and Breadth-First Searches
Problem. We have to traverse through a (di)graph to nd some kind of vertices or edges.
We assume that the (di)graph is connected and loopless. For disconnected graphs, we have to
go through the components separately. We ignore loops if the (di)graph has any.
Depth-First Search, DFS, has many uses. The procedure is a bit different for undirected
graphs and directed graphs. Therefore they will be treated separately.
Undirected Graphs
We choose a starting vertex r (root) to start the search. Then, we traverse an edge e = (r, v)
to go to the vertex v. At the same time, we direct e from r to v. Now, we say that the edge e
is examined and we call it a tree edge. The vertex r is called the father of v and we denote it
r = FATHER(v).
We continue the search. At a vertex x, there are two cases:
(1) If every edge incident to x has been examined, return to the father of x and continue the
process from FATHER(x). The vertex x is said to be completely scanned.
(2) If there exist some unexamined edges incident to x, then we choose one such edge e =
(x, y) and direct it from x to y. This edge is now said to be examined. We have two
subcases now:
(2.1) If y has not been visited before, then we traverse the edge (x, y), visit y and continue
the search from y. In this case, e is a tree edge and FATHER(y) = x.
(2.2) If y has been visited before, then we select some other unexamined edge incident to
x. In this case, the edge e is called a back edge.
Every time we come to a new vertex which has never been visited before, we give it a distinct
number. The number of the root is 1. We write
DFN(x) = running number of vertex x.
A complete DFS ends when we traverse back to the root and we have visited every vertex or
when we have found the desired edge/vertex.
DFS divides the edges of G into tree edges and back edges. Obviously, the tree edges form
a spanning tree of G, also known as a DFS tree. If we include the directions of the tree edges,
we get a directed DFS tree. DFS gives a direction to every edge in G. When we use these
directions, we get a digraph whose underlying subgraph is G. It has the DFS tree as a directed
spanning tree.
Example. For the graph
we start the DFS from a root in the upper left corner. The back edges are marked with two lines.
1
5
4
6
3
2
11
9
10
8
7
=
=
=
=
=
=
=
The corresponding DFS tree is
1
5
4
6
3
2
11
9
10
8
7
In the following, we denote,
K(x) =
0 if vertex x has not been visited

1 if vertex x has been visited
and TREE and BACK are set variables containing the directed tree edges and back edges.
Depth-First Search for Graphs:
1. Set TREE , BACK and i 1. For every vertex x of G, set FATHER(x) 0
and K(x) 0.
2. Choose a vertex r for which K(r) = 0 (this condition is needed only for disconnected
graphs, see step #6). Set DFN(r) i, K(r) 1 and u r.
3. If every edge incident to u has been examined, go to step #5. Otherwise, choose an edge
e = (u, v) that has not been examined.
4. We direct edge e from u to v and label it examined.
4.1 If K(v) = 0, then set i i +1, DFN(v) i, TREE TREEe, K(v) 1,
FATHER(v) u and u v. Go back to step #3.
4.2 If K(v) = 1, then set BACK BACK e and go back to step #3.
5. If FATHER(u) ,= 0, then set u FATHER(u) and go back to step #3.
6. (Only for disconnected graphs so that we can jump from one component to another.) If
there is a vertex r such that K(r) = 0, then set i i +1 and go back to step #2.
7. Stop.
We denote T as the DFS tree and

G as the directed graph obtained from the algorithm. T
is a directed spanning tree of

G. If there is a directed path from u to v in T, then we call u an
ancestor of v and v a descendent of u. Vertices u and v are related if one of them is an ancestor
of the other. In particular, if (u, v) is an edge of T, then u is the father of v and v is a son of u.
An edge (u, v) of G, where u and v are unrelated, is called a cross edge. However,
Statement. Cross edges do not exist.
Proof. Let u and v be two distinct vertices which are unrelated. Then, (by quasi-strong connec-
tivity) there are two vertices u
and v
such that
FATHER(u
) = FATHER(v
),
u
= u or u
is an ancestor of u and
v
= v or v
is an ancestor of v.
We examine the case where DFN(u
) < DFN(v
) (the other case is obviously symmetrical).

We label T
1
as the directed subtree of T whose root is u
and T
2
as the directed subtree of T
whose root is v
. Obviously, DFS goes through the vertices of T

2
only after u
is completely
scanned. Furthermore, u
is completely scanned only after all the vertices of T

1
are completely
scanned. Hence, it is impossible to have an edge (u, v).
Directed Graphs
Depth-rst search in a (connected and loopless) digraph G is similar to the case for undi-
rected graphs. The algorithm divides the arcs in G into four different classes. If the search
proceeds to an unexamined arc e = (x, y), then the four possible classes are:
(1) If y has not been visited, then e is a tree edge.
(2) If y has been visited, then there are three cases:
(2.1) y is a descendent of x in the subgraph induced by existing tree edges. Then, e is a
forward edge and DFN(y) > DFN(x).
(2.2) x is a descendent of y in the subgraph induced by the existing tree edges. Then, e is
a back edge and DFN(y) < DFN(x).
(2.3) x and y are not related by any of the existing tree edges. Then, e is a cross edge
and DFN(y) < DFN(x). (Note! It is impossible that DFN(y) > DFN(x). This is
proven in the same way as we did previously.)
The directed subgraph of G induced by tree edges is called the DFS forest (directed forest).
If DFN(y) > DFN(x) holds for the arc (x, y), then (x, y) is a tree edge or a forward edge.
During the search, it is easy to distinguish the two because (x, y) is a tree edge if y has not been
visited and it is a forward edge otherwise. If DFN(y) < DFN(x), then (x, y) is a back edge or
a cross edge. During the search, it is easy to distinguish the two because (x, y) is a cross edge
if y is completely scanned and it is a back edge otherwise.
In the following, K, FATHER, TREE and BACK are dened as previously. We also have
two new variables FORWARD and CROSS (their meanings are obvious) and
L(x) =
1 if x is completely scanned
0 otherwise.
Depth-First Search for Digraphs:
1. Set TREE , FORWARD , BACK , CROSS and i 1. For every
vertex x in G, set FATHER(x) 0, K(x) 0 and L(x) 0.
2. Choose a vertex r such that K(r) = 0 and set DFN(r) i, K(r) 1 and u r.
3. If every arc coming out of u has already been examined, then set L(u) 1 and go to
step #5. Otherwise, choose an unexamined arc e = (u, v).
4. Label the arc e examined.
4.1 If K(v) = 0, then set i i +1, DFN(v) i, TREE TREEe, K(v) 1,
FATHER(v) u and u v. Go to step #3.
4.2 If K(v) = 1 and DFN(v) > DFN(u), then set FORWARD FORWARD e
and go to step #3.
4.3 If K(v) = 1 and DFN(v) < DFN(u) and L(v) = 0, then set BACK BACK
e and go to step #3.
4.4 If K(v) = 1 and DFN(v) < DFN(u) and L(v) = 1, then set CROSS CROSS
e and go to step #3.
5. If FATHER(u) ,= 0, then set u FATHER(u) and to go step #3.
6. If there is a vertex r such that K(r) = 0, then set i i +1 and go to step #2.
7. Stop.
Example. DFS in the following digraph starts from a root in the upper left corner and proceeds
like this (back edges are marked with one line, cross edges are marked with two lines and
forward edges are marked with three lines):
1
2
3
4
5
6
7
8
9
10 11 12
13
The corresponding DFS forest is

1
2
3
4
5
6
7
8
9
10 11 12
13
Theorem 5.1. If a depth-rst search in a quasi-strongly connected digraph starts fromone of its
roots, then the DFS forest is a directed tree. In particular, the DFS forest of a strongly connected
digraph is a directed tree no matter where the search starts from.
Proof. Let us prove by contradiction and consider the counter hypothesis: The DFS forest T
resulted from a DFS in a quasi-strongly connected digraph G that began from root r is not a
directed tree.
Since T is a directed forest, the component T
1
of T which has the root r does not contain
some vertex v of G. On the other hand, there is a directed path from r to v. We choose the last
vertex u on this path which is in T
1
and the arc e = (u, w). Since the vertex w is not in T
1
,
the edge e is not a tree edge, a back edge nor a forward edge. Then, it must be a cross edge.
Because the search began at r, the vertex w has to be in T
1
(
).
Strongly connected digraphs are also quasi-strongly connected and any vertex can be chosen
as a root.
Breadth-rst search, BFS, is related to DFS. Let us consider a connected graph G.
Breadth-First Search for Graphs:
1. In the beginning, no vertex is labeled. Set i 0.
2. Choose a (unlabeled) starting vertex r (root) and label it with i.
3. Search the set J of vertices that are not labeled and are adjacent to some vertex labeled
with i.
4. If J ,= , then set i i +1. Label the vertices in J with i and go to step #3.
5. (Only for disconnected graphs so we can jump from one component to another.) If a
vertex is unlabeled, then set i 0 and go to step #2.
6. Stop.
BFS also produces a spanning tree, called the BFS tree, when we take the edges
(vertex labeled with i, unlabeled vertex)
while forming J. One such tree edge exists for each vertex in J. We obtain the directed BFS tree
by orienting the edges away from the labeled vertex to the unlabeled vertex. BFS as presented
above does not however orient every edge in the graph. Obviously, the label of a vertex is the
length of the shortest path from the root to it, in other words, the distance from the root.
Example. BFS in the graph we had in the previous example starts at a root in the upper left
corner and proceeds as follows. (Tree edges are marked with two cross lines.)
=
=
=
=
=
=
=
=
=
=
The corresponding BFS tree is
We obtain the directed BFS tree by orienting the branches away from the root.
BFS in a digraph G is very similar to what we just did.
Breadth-First Search for Digraphs:
1. In the beginning, no vertex is labeled. Set i 0.
2. Choose an unlabeled starting vertex r (root) and label it with i.
3. Search the set J of terminal vertices of arcs whose initial vertices have been labeled with
i but whose terminal vertices have not been labeled.
4. If J ,= , then set i i +1. Label the vertices in J with i and go to step #3.
5. If not all vertices have been labeled, then set i 0 and go to step #2.
6. Stop.
BFS in a digraph produces a BFS forest (directed forest) when we take the examined arcs
(vertices labeled withi, unlabeled vertices)
while forming J. One such tree edge exists for each vertex in J.
Remark. In addition, BFS can be modied to sort the arcs like DFS.
5.4 The Lightest Path: Dijkstras Algorithm
Problem. The edges of a (di)graph are given non-negative weights. The weight of a path is
the sum of the weights of the path traversed. We are to nd the lightest (directed) path in the
(di)graph from vertex u to vertex v (,= u) if the path exists (sometimes also called the shortest
path). We should state if such path does not exist.
Obviously, we can assume that we do not have any loops or parallel edges. Otherwise, we
simply remove the loops and choose the edge with the lowest weight out of the parallel edges.
From now on, we only consider directed graphs. Undirected graphs can be treated in the same
way by replacing an edge with two arcs in opposite directions with the same weight.
We denote (r, s) as the weight of the arc (r, s). Dijkstras Algorithm marks the vertices as
permanent or temporary vertices. The label of a vertex r is denoted (r) and we dene
(r) =
1 if the label is permanent

0 if the label is temporary.
A permanent label (r) expresses the weight of the lightest directed ur path. A temporary
label (r) gives an upper limit to this weight (can be ). Furthermore, we denote:
(r) =
the predecessor of vertex r on the lightest directed ur path if such a path exists
0 otherwise,
so we can construct the directed path with the lowest weight.
Dijkstras Algorithm:
1. Set (u) 0 and (u) 1. For all other vertices r, set (r) and (r) 0. For
all vertices r, we set (r) 0. Furthermore, set w u.
2. For every arc (w, r), where (r) = 0 and (r) > (w) +(w, r), set
(r) (w) +(w, r) and (r) w.
3. Find a vertex r
for which (r
) = 0, (r
) < and
(r
) = min
(r)=0
(r).
Set
(r
) 1 and w r
.
If there is no such vertex r
, a directed uv path does not exist and we stop.

4. If w ,= v, then go to step #2.
5. Stop.
We see that the algorithm is correct as follows. We denote (for every step):
V
1
= permanently labeled vertices
V
2
= temporarily labeled vertices.
(V
1
, V
2
is a cut with the completely scanned vertices on one side and other vertices on the
other side.)
Statement. The label (r) of the vertex r in V
1
is the weight of the lightest directed ur path
and (r) is the predecessor of r on such a path.
Proof. After step #2, the temporary label of r is always the weight of a directed ur path with
the lowest weight whose vertices are in V
1
except for r (= if there is no such path), and (r)
is a predecessor of r on this path (or = 0). This is because (two cases):
Before step #2, (r) = . The only new vertex in V
1
is now w so every possible
directed ur path has to visit w. If there is no such path, then the case is obvious ((r)
stays at and (r) stays at zero). Let us assume that we have the (lightest) directed
ur path that contains only vertices of V
1
and r as well. In particular, w is included. The
subpath from u to w has of course the lowest weight. We consider the vertex s ( V
1
)
which is the predecessor of r on the directed ur path. If s = w, then the case is clear. If
s ,= w, then s has been a w before, in which case (r) can not be = (step #2) (
).
Before step #2, (r) < . Then, (r) is the weight of the lightest directed ur path
whose vertices are in V
1
w except for r. The only new vertex in V
1
is w so every
possible lighter directed ur path has to visit w. If there is no such path, then the case
is obvious ((r) and (r) remain unchanged). Let us assume that we have a (lighter)
directed ur path that contains only vertices of V
1
and r as well. In particular, w is
included. The subpath from u to w has of course the lowest weight. We consider the
vertex s ( V
1
) which is the predecessor of r on the directed ur path. If s = w, then
the case is clear. If s ,= w, then s is in V
1
w. Since s has been a w before, there is
a lightest directed us path that does not contain w (otherwise, we should have chosen
r
in step #3 to be some predecessor of s on the directed uws path). Then, we get a

directed ur path with a lower weight that contains r and only vertices in V
1
w (
).
The permanent label is the weight we seek because of the minimization in step #3 and (r)
gives a predecessor of r as we claimed.
At the end of the algorithm, vertex v gets a permanent label or the process stops at step #3
(which means a directed uv path does not exist). The directed path with the lowest weight can
be obtained by starting from the vertex v and nding the predecessors by using the label .
If we replace step #4 by
4. Go to step #2.
and continue the process until it stops at step #3, we get
(w) =
0 if w = u
the weight of the lightest directed uw path if there is one
otherwise
and
(w) =
the predecessor of w on the lightest directed uw path if there is one and w ,= u

0 otherwise.
Remark. Dijkstras algorithm may fail if there are negative weights. These cases are investi-
gated in the next section.
5.5 The Lightest Path: Floyds Algorithm
Problem. We are to nd the lighest path from vertex u to vertex v (,= u) in a digraph or to show
that there is no such path when the arcs of the digraph have been assigned arbitrary weights.
Note that the weight of a directed path is the sum of the weights of the arcs traversed.
Obviously, we can assume there are no loops or parallel arcs. Otherwise, we simply remove the
loops and choose the arc with the lowest weight out of the parallel arcs. Floyds Algorithm only
works for digraphs. We write the weight of (x, y) as (x, y) and construct the weight matrix
W= (w
ij
) where
w
ij
=
(v
i
, v
j
) if there is an arc (v
i
, v
j
)
otherwise.
(Once again, V = v
1
, . . . , v
n
is the vertex set of the digraph.) Floyds Algorithm is similar
to Warshalls Algorithm. It only works if the digraph has no negative cycles, i.e. no directed
circuit in the digraph has a negative weight. In this case, the lightest directed path is the lightest
directed walk.
Floyds Algorithm constructs a sequence of matrices W
0
, W
1
, . . . , W
n
where W
0
= W
and
(W
k
)
ij
= weight of the lightest directed v
i
v
j
path,
where there are only vertices v
1
, . . . , v
k
on the path besides v
i
and v
j
( = if there is no such path).
Statement. When W
k
is computed from W
k1
by the formula
(W
k
)
st
= min(W
k1
)
st
, (W
k1
)
sk
+(W
k1
)
kt
,
then we get the previously mentioned sequence of weight matrices. If the digraph has negative
cycles, then the sequence is correct up to the point when one of the diagonal elements turns
negative for the rst time.
Proof. We use induction on k.
Induction Basis: k = 1. Since the digraph is loopless, the diagonal elements of W
0
can only
be and the lightest directed path (if there is one) is one of the following, and the statement is
obvious:
s t: v
s
v
t
v
s
v
1
v
t
s = t:
v
s
v
1
Induction Hypothesis: The statement is true for k < . ( 2)
Induction Statement: The statement is true for k = .
Induction Statement Proof: The diagonal elements of W
1
have to be nonnegative ( is
permitted) for us to get this k. Let us consider the case where s ,= t. (The case s = t is
analogous.) We have ve cases:
Vertex v
is on the lightest directed path but it is not v

s
or v
t
, i.e. ,= s, t. Let us
consider the directed subpath from v
s
to v
whose vertices other than v

s
and v
are in
v
1
, . . . , v
1
. Suppose the lightest directed v
s
v
path of this kind has common vertices

with the directed subpath from v
to v
t
other than v
itself, e.g. v
p
. The directed v
s
v
p
v
t
walk we get would be lighter than the original directed v
s
v
t
path. By removing
cycles, we would get a directed v
s
v
t
path that would be lighter and would only contain
v
s
as well as v
t
and the vertices v
1
, . . . , v
). (We have to remember that weights of

cycles are not negative!) Therefore, the directed subpath from v
s
to v
is the lightest di-

rected v
s
v
path which contains the vertices v

1
, . . . , v
1
as well as v
s
and v
. Similarly,
the directed subpath from v
to v
t
is the lightest directed v
v
t
path which contains the
vertices v
1
, . . . , v
1
as well as v
t
and v
. Now, we use the Induction Hypothesis:

(W
)
st
< (W
1
)
st
(check the special case (W
1
)
st
= ) and
(W
)
st
= (W
1
)
s
+(W
1
)
t
.
The directed v
s
v
t
path with the lowest weight exists and v
= v
s
. By the Induction
Hypothesis, (W
)
st
= (W
1
)
st
and
(W
1
)
s
+(W
1
)
t
= (W
1
)
+(W
1
)
t
(W
1
)
t
= (W
1
)
st
,
since (W
1
)
0 (possibly = ).
The directed v
s
v
t
path exists and v
= v
t
. By the Induction Hypothesis, (W
)
st
=
(W
1
)
st
and
(W
1
)
s
+(W
1
)
t
= (W
1
)
s
+(W
1
)
(W
1
)
s
= (W
1
)
st
,
since (W
1
)
0 (possibly = ).
The lightest directed v
s
v
t
path exists but v
is not on the path. Now, we construct the

lightest directed v
s
v
path and the lightest v
v
t
path which, in addition to the end ver-
tices, contain only vertices v
1
, . . . , v
1
, if it is possible. By combining these two paths,
we get a directed v
s
v
t
walk. By removing possible cycles from this walk, we get an as
light or even lighter v
s
v
t
path, which only contains vertices v
1
, . . . , v
as well as v
s
and
v
t
. (We have to remember that weights of cycles are not negative!) Therefore, this is a
case where
(W
1
)
s
+(W
1
)
t
(W
1
)
st
and the equation in the statement gives the right result. If there is no directed v
s
v
path
or v
v
t
path, then it is obvious.
The lightest directed v
s
v
t
path does not exist. Then, (W
)
st
= and (W
1
)
st
= .
On the other hand, at least one of the elements (W
1
)
s
or (W
1
)
t
is = because
otherwise we would get a directed v
s
v
t
path by combining the v
s
v
path with the v
v
t
path as well as removing all possible cycles (
).
Floyds Algorithm also constructs another sequence of matrices Z
0
, . . . , Z
n
in which we
store the lightest directed paths in the following form
(Z
k
)
ij
=
where v
is the vertex following v

i
on the lightest directed
v
i
v
j
path containing only vertices v
i
and v
j
as well as v
1
, . . . , v
k
(if such a path exists)
0 otherwise.
Obviously,
(Z
0
)
ij
=
j if (W)
ij
,=
0 otherwise.
The matrix Z
k
(k 1) of the sequence can be obtained from the matrix Z
k1
by
(Z
k
)
ij
=
(Z
k1
)
ik
if (W
k1
)
ik
+(W
k1
)
kj
< (W
k1
)
ij
(Z
k1
)
ij
otherwise,
so the sequence can be constructed with the sequence W
0
, W
1
, . . . , W
n
at the same time.
Finally, Floyds Algorithm is presented in the following pseudocode. We have added a part
to test if there are negative elements on the diagonal and the construction of the Z
0
, . . . , Z
n
sequence of matrices.
procedure Floyd
begin
W:= W
0
k := 0
for i := 1 to n do
for j := 1 to n do
if (W)
ij
= then
(Z)
ij
:= 0
else
(Z)
ij
:= j
od
od
while k < n and Test(W) do
Iteration(W, Z, k)
od
end
subprocedure Test(W)
begin
for i := 1 to n do
if (W)
ii
< 0 then
return FALSE
od
return TRUE
end
subprocedure Iteration(W, Z, k)
begin
k := k +1
for i := 1 to n do
for j := 1 to n do
if (W)
ik
+(W)
kj
< (W)
ij
then
(W)
ij
:= (W)
ik
+(W)
kj
(Z)
ij
:= (Z)
ik
od
od
end
5.6 The Lightest Spanning Tree: Kruskals and Prims Algo-
rithms
Problem. We have to nd the spanning tree with the lowest weight of a connected graph if the
edges of the graph have been weighted arbitrarily and the weight of a tree is the sum of all the
weights of the branches.
Obviously, we can assume that the graph G = (V, E) is nontrivial and simple. Otherwise, we
simply remove the loops and choose the edge with the lowest weight out of the parallel edges.
We denote the weight of the edge e as (e) and the weight of the spanning tree T as (T).
As usual, we write the number of vertices as n, number of edges as m, V = v
1
, . . . , v
n
and
E = e
1
, . . . , e
m
.
The distance between two spanning trees T
1
and T
2
of G is
n 1 #(T
1
T
2
) =
def.
d(T
1
, T
2
),
where #(T
1
T
2
) is the number of edges in the intersection of T
1
and T
2
. Obviously, d(T
1
, T
2
) =
0 if and only if T
1
= T
2
. If d(T
1
, T
2
) = 1, then T
1
and T
2
are neighboring trees.
The spanning tree T of G is cut minimal if the weights of the edges of the fundamental
cut set determined by the branch b are (b) for every branch b. Similarly, the spanning tree
T is circuit minimal if the edges of the fundamental circuits are (c) for every link c in
the cospanning tree T
. The spanning tree T is locally minimal if (T) (T
) for every
neighboring tree T
of T.
Lemma. The following three conditions are equivalent for the spanning tree T:
(i) T is cut minimal.
(ii) T is circuit minimal.
(iii) T is locally minimal.
Proof. (i)(ii): Let us assume T is cut minimal and let us consider a fundamental circuit C of
G corresponding to the link c of the cospanning tree T
. Other than c, the branches in C are

branches of T. Every such branch b denes a fundamental cut set of T, which also contains c
(Theorem 2.7). Hence (b) (c).
(ii)(iii): Let us assume that T is circuit minimal and let us consider a neighboring tree T
of T. T
has (exactly one) branch e
which is not in T, i.e. e
is a link of T
. We examine the
fundamental circuit C dened by e
. Not all edges of C are in T
. We choose an edge e in C
that is not in T
. Then, e is a branch of T (actually the only branch of T that is not in T
). Now,
we remove e out of T and add e
to T. The result has to be T
. Because of circuit minimality,

(e
) (e), i.e. (T
) (T).
(iii)(i): We consider the locally minimal spanning tree T. We take an arbitrary branch
b from T corresponding to a fundamental cut set I and an arbitrary link c ,= b in I. Then, b
belongs to the fundamental circuit of T dened by c (Theorem 2.8). By removing the branch
b from T and adding the edge c to T, we get the neighboring tree T
of T. Because of local
minimality, (T) (T
), i.e. (c) (b).

The spanning tree T is minimal if it has the lowest possible weight.
Theorem 5.2. The following three conditions are equivalent for the spanning tree T:
(i) T is cut minimal.
(ii) T is circuit minimal.
(iii) T is minimal.
Proof. By the lemma above, (i) and (ii) are equivalent. A minimal spanning tree is obviously
locally minimal. Thus, it sufces to prove that a cut minimal spanning tree is also minimal.
We will prove by contradiction and consider the counter hypothesis: There is a cut minimal
spanning tree T which is not minimal. Let us consider the minimal spanning tree T
and choose
T and T
so that the distance d(T, T
) is as small as possible. By the lemma, d(T, T
) > 1.
T has a branch e which is not in T
, i.e. it is a link of (T
. We label the fundamental

cut set of T dened by e as I and the fundamental circuit of T
dened by e as C. In the
intersection I C, there are also other edges besides e (Theorem 2.6). We choose such an edge
e
. Then, e
is a link of T
and a branch of T
. Since T is cut minimal, (e
) (e). Since
T
is (circuit) minimal, (e
) (e). Therefore, (e
) = (e). By removing e
from T
and
adding e to T
, we get a minimal spanning tree T
which has the same weight as T
. However,
d(T, T
) < d(T, T
).
In Kruskals Algorithm, the edges of the graph G(and their weights) are listed as e
1
, . . . , e
m
.
The algorithm constructs a circuit minimal spanning tree by going through the list to take some
edges to form the tree. This is especially effective if the edges are sorted in ascending order by
weight.
In the dual form of Kruskals Algorithm, we construct a cut minimal spanning tree by going
through the list of edges to take some edges to form the cospanning tree. Once again, this is
especially effective if the edges are sorted in descending order by weight.
In all, we get four different versions of Kruskals Algorithm. (We have to remember that the
subgraph induced by the edge set A is written as A.)
Kruskals Algorithm No. 1
Here we assume that the edges are given in ascending order by weight.
1. Set k 1 and A .
2. If e
k
does not form a circuit with the edges in A, then set A A e
k
as well as
k k +1 and go to step #4.
3. If e
k
forms a circuit with the edges in A, then set k k +1 and go to step #4.
4. If (V, A) is not a tree, then go to step #2. Otherwise stop and output the spanning tree
T = A.
Whenever we leave out an edge from A (step #3), its end vertices are already connected in A.
Thus, the vertices of G are connected in T as they are in G. Since T is obviously circuitless
(step #3), it is also a spanning tree of G. At each stage, the branches of the fundamental circuit
dened by the link belonging to T
(step #3) are predecessors of that link in the list. Hence, T

is circuit minimal and thus minimal.
Remark. In every step, the branches and links are permanent. We do not have to know the
edges beforehand as long as we process them one by one in ascending order. The rank of the
graph (number of branches in a spanning tree) is then required beforehand so we know when to
stop.
Here we assume the edges are given in an arbitrary order.
1. Set k 1 and A .
2. If A e
k
contains no circuits, then set A A e
k
as well as k k + 1 and go
to step #4.
3. If Ae
k
contains a circuit C, then choose the edge with the largest weight e in C (if
there are more than one, take any), set A (A e
k
) e as well as k k +1 and
go to step #4.
4. If k m, then go to step #2. Otherwise, stop and output the spanning tree T = A.
Whenever we leave out an edge from A (step #3), its end vertices are already connected in A.
Thus, the vertices of G are connected in T as they are in G. Since T is obviously circuitless
(step #3), it is a spanning tree of G.
We see that T is circuit minimal (and minimal) by the following logic. During the whole
process, A is a forest by step #4. In addition, if u and w are connected in A at some point,
then they are also connected afterwards. The uw path in A is unique but it can change to
another path later in step #3. Nevertheless, whenever this change occurs, the maximum value
of the weights of the edges of the path can not increase anymore. Every link c of T
has been
removed from A in step #3. Then, the weight of c is at least as large as the weights of the other
edges in C, After we have gone through step #3, the only connected end vertices of c in A
have to go through the remaining edges of C. The nal connection between the end vertices of
c in T goes through the edges of the fundamental circuit dened by c. Therefore, the weights
of the edges of this fundamental circuit are (c).
Remark. In each step, the links (e in step #3) are permanent and the branches are not. We do
not have to know the edges beforehand as long as we process them one by one. However, we
need to know the nullity of the graph (number of links in a cospanning tree) so that we know
when to stop. The algorithmcan also be used to update a minimal spanning tree if we add edges
to the graph or decrease their weight.
Here we assume the edges are given in descending order by weight.
1. Set A E and k 1.
2. If (V, A e
k
) is connected, then set A A e
k
as well as k k + 1 and go to
step #4.
3. If (V, Ae
k
) is disconnected, then set k k +1 and go to step #4.
4. If (V, A) is not a tree, then we go to step #2. Otherwise we stop and output the spanning
tree T = (V, A).
T is obviously connected because (V, A) is connected everytime we go to step #4. On the other
hand, T is circuitless because if the circuit C is in T and the edge c is in the circuit, then c is
removed from A in step #2 when e
k
= c (
). Thus, T is a spanning tree of G. In each step, the

links of the fundamental cut set dened by the branch belonging to T (step #3) are predecessors
of that branch in the list. Hence, T is cut minimal and it is thus minimal.
Remark. In each step, the branches and links are permanent. We have to know the edges
beforehand. On the other hand, we do not have to know their weights as long as we get them
one by one in descending order.
Here we assume the edges are given in an arbitrary order.
1. Set A E and k 1.
2. If (V, A e
k
) is connected, then set A A e
k
as well as k k + 1 and go to
step #4.
3. If (V, A e
k
) is disconnected, then it has two components. The corresponding vertex
sets form a cut V
1
, V
2
. We interpret it as an edge set and choose the edge e with the
lowest weight in V
1
, V
2
(if there are more than one, take any). Set A (Ae
k
)e
as well as k k +1 and go to step #4.
4. If k m, then go to step #2. Otherwise stop and output the spanning tree T = (V, A).
T is obviously connected because (V, A) is connected everytime we go to step #4. (Take note
that the connectivity is preserved everytime we go through step #3.) On the other hand, T is
circuitless. If a circuit C of G ends up in T and c is the edge of C, which is rst in the list,
then c must be removed from A in step #2 when e
k
= c. (Note that the edge removed from the
circuit rst can not be removed in step #3.) If c comes back later (in step #3), then it forms a
cut set of (V, A) by itself in which case some other edge of C has been removed. By continuing
this process, we see that all the edges of C can not be in A in the end (
). Therefore, T is a
spanning tree of G.
In addition, T is cut minimal and minimal because every branch b of T comes in in step
#3. The links of the fundamental cut set dened by b are either edges of the cut V
1
, V
2
which
is examined at that point or they are links of the cuts we examined later in step #3. Whenever
an edge of this kind gets removed later in step #3, it is always compensated by edges that are
heavier in weight than b. Those heavier edges are in the cut V
1
, V
2
which is examined at that
time. Therefore, the weights of the fundamental cut set dened by b are (b).
Remark. In each step, the branches (e in step #3) are permanent and the links are not. We
have to know the edges beforehand. We do not have to know the weights beforehand as long as
we process them one by one. This algorithm can also be used for updating a minimal spanning
tree if we remove edges from a graph or if we increase the weights of edges.
Prims Algorithm
In Prims Algorithm (also known as Jarniks Algorithm), we use the all-vertex incidence ma-
trix of G. If we label the set of edges incident on vertex v as (v), then we can get a list
(v
1
), . . . , (v
n
), i.e. the cuts dened by the vertices (interpreted as edge sets). In addition, we
assign weights to the vertices.
The algorithm works in the same way as Dijkstras Algorithm by constructing the spanning
tree branch by branch. The variables are A (set of branches of the spanning tree we have at the
time), B (set of vertices of the spanning tree we have at the time) and I (the cut interpreted as
an edge set from which we choose the next branch).
Prims Algorithm (First Version):
1. Choose a starting vertex r and set A , B r as well as I (r).
2. Choose the lightest edge e from I (if there are more than one, choose any). Take the end
vertex v of e that is not in B. Set A A e, B B v as well as I I (v)
and go to step #3. (Remember that denotes the symmetric difference operation between
two sets, see page 12.)
3. If B ,= V , then go to step #2. Otherwise, stop and output the spanning tree T = (B, A) =
A.
Since the edge e was chosen from a cut, T is circuitless. On the other hand, because there is
a path from r to every other vertex, T has every vertex of G and it is connected. T is thus a
spanning tree. It is also minimal because
Statement. During the whole process, (B, A) is a subtree of some minimal spanning tree of G.
Proof. We use induction on , the number of vertices in B.
Induction Basis: = 1. The case is obvious because (B, A) is trivial.
Induction Hypothesis: The statement is true for = k 1. (k 2)
Induction Statement: The statement is true for = k.
Induction Statement Proof: In step #2, we can write A = A
e, where e I
and
B = B
v. (B
, A
) is a subtree of some minimal spanning tree T

min
from the induction
hypothesis. If e belongs to T
min
, then the case is clear. Otherwise, there is a fundamental circuit
C in T
min
+ e and there is another edge e
of I
in C (Theorem 2.6). Then, (e
) (e) and
(T
min
+e) e
is also a minimal spanning tree and (B, A) is its subtree (because T

min
is circuit
minimal and (e
) (e)).
Often, we use one or two additional labels for the vertices to make Prims Algorithm easier.
In the next version of the algorithm, we will use two labels (v) and (v), which are used to
perform step #2 more effectively. The values of are weights (up to ) and the values of are
edges (or = 0). Otherwise, the algorithms works in the same way as before.
Prims Algorithm (Second Version):
1. Choose a starting vertex r and set (r) 0. For every other vertex v, set (v) .
For every vertex v, set (v) 0 as well as A and B .
2. Choose a vertex u / B for which
(u) = min
v/ B
(v).
Set B B u. If (u) ,= 0, then set A A (u).
3. Go through all the edges e = (u, v) where v / B. If (e) < (v), then set (v) (e)
and (v) e.
4. If B ,= V , then go to step #2. Otherwise, stop and output the spanning tree T = (B, A) =
A.
5.7 The Lightest Hamiltonian Circuit (Travelling Salesmans
Problem): The Annealing Algorithm and the KarpHeld
Heuristics
Problem. If it is possible, we are to nd the Hamiltonian circuit with the lowest weight. A
Hamiltonian circuit visits all the vertices of a graph. As usual, the weights of the edges have
been assigned and the weight of a (directed) circuit is the sum of the weights of the edges
traversed.
Obviously, we can assume that the graph is nontrivial, connected (otherwise it would not be
possible to get a Hamiltonian circuit) and simple. If not, then we simply remove all the loops
and choose the edge with the lowest weight out of the parallel edges. As usual, we denote n as
the number of vertices, m as the number of edges, V = v
1
, . . . , v
n
and E = e
1
, . . . , e
m
.
We label the weight of an edge e = (v
i
, v
j
) as (e) = (v
i
, v
j
) and the weight of a Hamiltonian
circuit H as (H). We agree that the rst vertex of a Hamiltonian circuit is v
1
.
The same problem exists for directed graphs in which case we are looking for the directed
Hamiltonian circuit with the lowest weight (known as the Unsymmetric Travelling Salesmans
Problem).
The Travelling Salesmans Problem (TSP)
1
is an AT-complete problem, read e.g. MEHL-
HORN for more information. Actually, even deciding the existence of a Hamiltonian circuit is
an AT-complete problem. Solving a small TSP takes a lot of time and larger problems take so
much time that it is almost impossible to obtain accurate solutions. Therefore, many stochastic
and approximation methods are used in practice. Then, we have to accept the possibility of
inaccurate outcomes or even the lack of results.
1
The name Travelling Salesmans Problem comes from an interpretation where the vertices of a graph are
cities and the weights of the edges between the cities are travelling times. The salesman needs to visit every city
in the shortest amount of time.
The Annealing Algorithm
The annealing algorithms or thermodynamic algorithms have the following common features:
(A) The system in question is always in some state s. The set of all states S is nite and
known. In the TSP, a state is a Hamiltonian circuit.
(B) Each state s has a response f(s), which can be calculated in a timely fashion from the
state. Our goal is to nd a state whose response is near the minimum/maximum value.
The response of a state of a TSP is the weight of a Hamiltonian circuit.
(C) There is a procedure A
k
which is used to move fromstate s to state A
k
(s). k is a parameter
of the procedure which belongs to the set K. K can change during the procedure. The
purpose is to move to certain states near the state s which are dened by the parameter
k. By repeating the procedure with proper values of k, we should be able to move from
any state to any other state. (In some cases, we can omit this last part.)
(D) Every time we move from one state to another, we should be able to choose the param-
eter k quickly and randomly from K. In particular, the set K itself should be easily
computable.
(E) We should be able to quickly perform the procedure A
k
given a value of k.
(F) We should be able to nd a starting state s
0
. For the TSP, the starting state is a Hamiltonian
circuit.
The algorithm is as follows:
The Annealing Algorithm:
1. Choose the starting state s
0
, the initial temperature T
0
and set s s
0
as well as T T
0
.
2. When we are in the state s, we randomly choose a parameter k K and compute s
=
A
k
(s).
3. If f(s
) f(s), then set s s
and go to step #5.

4. If f(s
) > f(s), then generate a random number r in the interval [0, 1). If r e
f(s)f(s
)
T
,
then set s s
. Thus, we accept a worse state with probability e

f(s)f(s
)
T
. Note that the
greater the temperature T, the greater the probability that we go uphill.
5. If we have gone through a maximum total number of iterations, then we stop and output
s. Otherwise, if we have gone through sufciently many iterations of the procedure using
temperature T, then we lower T by some rule and go to step #2.
Remark. The distribution of the probability p
s
= e
f(s)f(s
)
T
used in step #4 is (apart from
normalizing) a so-called maximum entropy distribution with the following condition on the
expected value:
=A
k
(s)
kK
f(s
)>f(s)
p
s
f(s
) =
where depends on T and s. The distribution is also called a Boltzman distribution and it is
analogous to the distribution of the same name in Statistical Mechanics. Refer to courses in
Physics and Information Theory for more information.
At rst, we wait until the uctuation in the states settles to a certain equilibrium (using the
response f(s)). After that, we lower the value of T a bit and wait again for the equilibrium.
Then, we lower T again and so on. We continue this until the change in values of f(s) is
sufciently small or if we have ran out of time.
The operation A
k
and the set K of the neighboring states depend on the problem. The state
structure and the response function also depend on the problem. For the TSP, we still need to
assign A
k
and K for every situation. For this purpose, we take another parameter j and set
j 2. In step #2, we update j in the following way:
j
j +1 if j < n
2 otherwise.
(Another way of choosing j in step #2 would be to choose it randomly out of 2, . . . , n.)
Furthermore, we choose
K = 2, . . . , n j.
A
k
is dened by the following operation (known as the reversal):
If k > j, then we reverse the order of the vertices v
i
j
, . . . , v
i
k
on the corresponding
subpath in the current Hamiltonian circuit
s : v
1
, v
i
2
, . . . , v
in
, v
1
.
If k < j, then we reverse the order of the vertices v
i
k
, . . . , v
i
j
on the corresponding
subpath in the current Hamiltonian circuit
s : v
1
, v
i
2
, . . . , v
in
, v
1
.
We add the missing edges to the graph with very large weights so that we get a complete
graph and we will not have to worry about the existence of a Hamiltonian circuit in the rst
place. If we still do not get a Hamiltonian circuit in the end without those added edges, then
there is not a Hamiltonian circuit.
The starting temperature T
0
should be much larger than the values of [f(s
) f(s)[ which
guarantees that we can in principle move to any state (annealing) in the earlier stages of the
algorithm. After that, we lower the temperature applying some rule, for example a 10% change.
The annealing algorithm also works for the unsymmetric TSP with obvious changes.
KarpHeld Heuristics
In the KarpHeld Heuristics, we do not directly look for a Hamiltonian circuit but look for a
similar subgraph, known as a spanning 1-tree
2
. The process does not work for the unsymmetric
TSP. The spanning 1-tree S
v
corresponding to the vertex v (known as the reference vertex) is a
subgraph of G that satises the following conditions:
(a) S
v
is connected and contains every vertex of G.
(b) S
v
contains exactly one circuit C and the vertex v belongs to C.
(c) S
v
has exactly two edges incident on v.
2
Not to be confused with the 1-tree on p. 23!
Clearly, a Hamiltonian circuit is a spanning 1-tree corresponding to any of the vertices. The
weight of the spanning 1-tree S
v
is the sum of of the weights of all its edges, denoted (S
v
). S
v
is minimal if it has the lowest possible weight.
Statement. S
v
is minimal if and only if
(i) S
v
v is a minimal spanning tree of Gv, and
(ii) the two edges of S
v
incident on v are the two lightest edges of G out of all the edges
incident on v.
Proof. Let S
v
be a minimal spanning 1-tree. Let e and e
be the two edges in S

v
incident on
v. Then, S
v
v is a spanning tree of G v because removing v destroys the circuit but the
connections remain unsevered. If S
v
v is not a minimal spanning tree of G v then there is
a lighter spanning tree T of Gv. By adding the vertex v and the edges e and e
to T, we get a
spanning 1-tree corresponding to vertex v which is lighter than S
v
(
). Therefore, (i) is true.

Obviously, (ii) is true (because otherwise we would get a lighter spanning 1-tree by replacing e
and e
with the two lightest edges in G incident on v).

Let us assume that (i) and (ii) are true. If S
v
is not minimal, then there is a lighter minimal
spanning 1-tree S
v
corresponding to v. Because S
v
also satises (ii), the two edges incident on
v are the same (or at least they have the same weight) in S
v
and S
v
. Thus, S
v
v is lighter than
S
v
v (
).
It follows from the statement that any algorithm that nds the minimum spanning tree also
works for nding the minimumspanning 1-tree with minor modications. Especially, Kruskals
and Prims Algorithms are applicable.
In the KarpHeld Heuristics, we also use weights of vertices, denoted (v). With these, we
can dene the virtual weight of an edge as
(v
i
, v
j
) = (v
i
, v
j
) +(v
i
) +(v
j
).
With the concept of virtual weights, we get the virtual weight of a spanning 1-tree S
v
(we label
the edge set of S
v
as A):
(S
v
) =
(v
i
,v
j
)A
(v
i
, v
j
) =
(v
i
,v
j
)A
(v
i
, v
j
) +
(v
i
,v
j
)A
((v
i
) +(v
j
))
= (S
v
) +
(v
i
,v
j
)A
((v
i
) +(v
j
)).
Now we denote the degree of the vertex u in S
v
as d
Sv
(u). Then,
(v
i
,v
j
)A
((v
i
) +(v
j
)) =
n
i=1
(v
i
)d
Sv
(v
i
)
and
(S
v
) = (S
v
) +
n
i=1
(v
i
)d
Sv
(v
i
).
In particular, if we have a Hamiltonian circuit H (a special spanning 1-tree), then
d
H
(v
1
) = = d
H
(v
n
) = 2
and
(H) = (H) + 2
n
i=1
(v
i
)

Does not depend on H!
.
Minimization of the Hamiltonian circuits using virtual weights yields the same minimal circuit
than obtained by using real weights. In general though, if we use virtual weights to search for
spanning 1-trees, then we get results different from the spanning 1-trees obtained by using real
weights.
From now on, we only consider the spanning 1-tree corresponding to vertex v
1
and we leave
out the subscript. This is not a limitation of any kind on the Hamiltonian circuits although it
might be a good idea to change the reference vertex every now and then. We assume that H
min
is a minimal Hamiltonian circuit and S
is the minimal spanning 1-tree obtained from using

virtual weights (which of course corresponds to v
1
). Then,
(H
min
)
(S
).
In addition,
(H
min
) = (H
min
) +2
n
i=1
(v
i
)
and
(S
) = (S
) +
n
i=1
(v
i
)d
S
(v
i
).
Thus,
(H
min
) =
(H
min
) 2
n
i=1
(v
i
)
(S
) 2
n
i=1
(v
i
)
= (S
) +
n
i=1
(v
i
)(d
S
(v
i
) 2),
from which we get a lower limit on (H
min
).
The idea of the KarpHeld Heuristics is to guide the degrees of the vertices in S
to the value
2 by changing the weights of the vertices. If we succeed, then we get a minimal Hamiltonian
circuit. In all cases, we get a lower limit on the weights (H) of the (possible) Hamiltonian
circuits by using the calculation above. (Note that d
S
(v
1
) is always = 2 if S
is the spanning
1-tree corresponding to v
1
.)
The KarpHeld Heuristics:
1. Set (v) 0 for every vertex v.
2. Set
(u, v) (u, v) +(u) +(v) for each edge (u, v).

3. Find the minimal spanning 1-tree S
using virtual weights
(u, v). If we fail to nd this

kind of spanning 1-tree, then there is no Hamiltonian circuit and we can stop.
4. If S
is a circuit, then output the minimal Hamiltonian circuit H = S
and stop.
5. If S
is not a circuit and the lower limit calculated from S
increased during the last K

iterations, then set (v) (v) +d
S
(v) 2 for every vertex v and go to step #2. (K is
a xed upper bound on the number of iterations.)
6. If the lower limit calculated from S
has not increased during the last K iterations, then

output that lower limit and stop.
This procedure does not always produce a minimal Hamiltonian circuit even if there exists
one. In practice, it often produces either a minimal Hamiltonian circuit or a good lower limit on
the weight of it. Getting a number for the lower limit does not, however, guarantee the existence
of a Hamiltonian circuit in the graph!
KarpHeld Heuristics has many steps where we have to choose between different options
(such as the reference vertex and the spanning 1-tree). We can not go through every possibility
so we must choose randomly. Then, we have a Las Vegas algorithm or stochastic algorithm.
5.8 Maximum Matching in Bipartite Graphs: The Hungar-
ian Algorithm
A matching in the graph G = (V, E) is a set of edges S E none of which are adjacent to each
other. A matching is a maximum matching if it has the greatest possible number of edges. The
end vertex of an edge in a matching is matched.
Problem. We want to nd the maximum matching in a bipartite graph.
An alternating path of a matching S is a path that satises the following conditions:
(1) The rst vertex on the path is not matched, and
(2) every second edge is in the matching and the remaining edges are not in the matching.
Note that the rst edge in an alternating path is not in the matching. In addition, if the last vertex
of an alternating path is not matched, then this path is an augmenting path of S. A matching
without augmenting paths is called a maximal matching.
Example. For the bipartite graph
v
1
v
2
v
3
v
4
v
5
w
1
w
2
w
3
w
4
w
6
w
5
G: S:
v
1
v
2
v
3
v
4
v
5
w
1
w
2
w
3
w
4
w
6
w
5
one augmenting path of the matching S = (v
1
, w
3
), (v
3
, w
2
), (v
4
, w
6
), (v
5
, w
5
) is the path
where the vertices are v
2
, w
2
, v
3
, w
6
, v
4
, w
1
.
v
1
v
2
v
3
v
4
v
5
w
1
w
2
w
3
w
4
w
6
w
5
We can augment a matching S using its augmenting path p as follows:
1. We remove the edges of S in p, and
2. We add the edges in p which are not in S.
The new edge set is obviously a matching. Note that the number of edges in S on an augmenting
path is one fewer than the number of the remaining edges. Therefore, the number of edges in
a matching increases by one after the augmenting operation. It is not possible to augment a
maximal matching.
Example. (Continuing from the previous example) By using the given augmenting path from
the matching S, we get a new maximal matching S
1
= (v
1
, w
3
), (v
2
, w
2
), (v
3
, w
6
), (v
4
, w
1
),
(v
5
, w
5
).
w
1
w
2
w
3
w
4
w
6
w
5
v
1
v
2
v
3
v
4
v
5
In the Hungarian Algorithm, we systematically search for augmenting paths until we get
a maximal matching. After that, it sufces to prove that a maximal matching is a maximum
matching. From now on, we only consider bipartite graphs because the algorithm is then much
simpler. We search for augmenting paths by constructing an alternating tree of a matching S
which is a subtree of G such that
(1) a vertex r (the root of the tree) is unmatched,
(2) every second edge on each path out from r is in S and the remaining edges are not in S,
and
(3) either there is an augmenting path out from r or we can not add any more edges to S.
An alternating tree is an augmenting tree if it has an augmenting path. Otherwise, it is a Hun-
garian tree. Every augmenting path is obviously an augmenting tree by itself. Note that the
only unmatched vertex of a Hungarian tree is the root.
Example. (Continuing from the previous example) Two alternating trees of the matching S are
(the root is circled):
w
1
w
2
w
4
w
6
v
2
v
3
v
4
v
2
v
3
v
4
w
2
w
4
w
6
Both of them are augmenting trees. An alternating tree of the matching S
1
(the root is w
4
) is the
Hungarian tree
v
1
v
4
w
1
w
3
w
4
Augmenting and Hungarian trees are not unique. We can have many different trees depend-
ing on the order we take the edges for constructing the trees even though the roots are the same.
On the other hand,
Statement. If a matching in a bipartite graph G has a Hungarian tree, then it does not have an
augmenting tree with the same root.
Proof. Let us prove by contradiction and consider the counter hypothesis: A matching S has a
Hungarian tree U and an augmenting tree T with the same root r. We get an augmenting path
in the augmenting tree
p : r = v
0
, e
1
, v
1
, e
2
, . . . , e
k
, v
k
.
We choose the last vertex v
i
which is in U from the path p (at least r = v
0
is in U). Since
v
k
is not in U, i < k. Furthermore, the edge e
i+1
is not in the matching nor in U (otherwise
v
i+1
would also be in U). On the other hand, since e
i+1
is not in U, v
i+1
has to be an end
vertex of another edge in U (
) because the only reason why the edge e

i+1
is not put into U
while constructing U is that the other end vertex v
i+1
of e
i+1
is already in U. Note how the
bipartiteness of the G comes in: If the cut in G that results in the bipartition is V
1
, V
2
, then the
vertices of U and p alternate between V
1
and V
2
. Therefore, the length of the rv
i
path is even
in p and U.
Constructing an alternating tree from a root always leads to a Hungarian tree or an augmenting
tree but not both. The order of the edges taken does not matter. (This is not the case for general
graphs.)
For the bipartite graph G = (V, E), the Hungarian Algorithm is as follows. The cut that
yields the bipartition is V
1
, V
2
.
The Hungarian Algorithm:
1. Set S . (We can also use some other initial matching.)
2. If every vertex in V
1
or in V
2
is matched in S, then S is a maximum matching and we
stop.
3. If there are unmatched vertices in S of V
1
, then go through them in some order construct-
ing alternating trees (the method of construction is not important as we claimed). If there
is an augmenting tree, then augmenting the matching S by using the augmenting path we
have another matching S
1
. Set S S
1
and go to #2.
4. If all the alternating trees that have unmatched vertices in V
1
as roots are Hungarian, S is
maximal and we stop.
Theorem 5.3. A maximal matching in a bipartite graph is a maximum matching.
Proof. Let us prove by contradiction and consider the counter hypothesis: A maximal matching
S in the bipartite graph G = (V, E) is not a maximum matching. Then, there are more edges in
the maximum matching S
max
than in S and in V
1
there are more vertices matched in S
max
than
in S. We choose an arbitrary vertex v V
1
, which is matched in S
max
but not in S. Then, we
have a path
p : v = v
0
, e
1
, v
1
, e
2
, . . . , e
k
, v
k
= w,
whose edges are alternating between S
max
and S, i.e. e
1
is in S
max
and e
2
is in S and so on.
We choose the longest such path p. Because p is obviously an alternating path of S, it has
even length, i.e. e
k
is an edge of S. (Otherwise, p would be an augmenting path of S which is
impossible because S is maximal.) Thus, w is matched in S but not matched in S
max
(because
the path p can not be continued).
Hence, every vertex v V
1
which is matched in S
max
but not in S corresponds to a vertex
w V
1
, which is matched in S but not in S
max
. Now, every path that ends at w must start
from the vertex v if the starting vertex is matched in S
max
but not in S. The last edge of such a
path has to be e
k
(the only edge in S incident on w) and the second to the last vertex has to be
v
k1
. Furthermore, the second to the last edge of this path has to be e
k1
(the only edge of S
max
incident on v
k1
) and the third to the last vertex has to be v
k2
, and so on.
However, there are then in V
1
at least as many vertices w that are matched in S but not in
S
max
as there are vertices v that are matched in S
max
but not in S (
).
Corollary. The Hungarian algorithm produces a maximum matching in a bipartite graph.
A matching is perfect if it matches every vertex of a graph. Thus, a graph with an odd
number of vertices can not have a perfect matching. Let us consider the graph G = (V, E) and
denote (v) = adjacent vertices of v as well as (A) =
vA
(v) for the vertex set A V .
Let us denote by #(X) the number of elements in the set X (the cardinality of the set). With
these notions, we can present the following famous characterization:
Theorem 5.4. (Halls Theorem or Marriage Theorem) A bipartite graph G whose bipar-
tition cut is V
1
, V
2
has a perfect matching if and only if every vertex set A V
1
and B V
2
satises the conditions #(A) #((A)) and #(B) #((B)).
Proof. If a perfect matching exists, then obviously #(A) #((A)) and #(B) #((B))
hold for all sets of vertices A V
1
and B V
2
. (Otherwise, we can not nd a pair for every
vertex in A or B in the matching.)
Let us assume that for all sets of vertices A V
1
and B V
2
, #(A) #((A)) and
#(B) #((B)). Let S be a maximum matching in G. We will prove by contradiction
and consider the counter hypothesis: S is not perfect. We choose a vertex v which is not
matched in S. Let us examine the case where v V
1
(the other case where v V
2
is obviously
symmetrical). The contradiction is apparent if v is an isolated vertex so we can move to the case
where v is an end vertex of an edge. The alternating tree with the root v is then nontrivial and
since the matching is also maximal, this tree is Hungarian. We choose such a Hungarian tree U.
We label the set of vertices of V
1
(resp. V
2
) in U by A (resp. by B). Because of the construction
of U, B = (A). On the other hand, the vertices of A and B in U are pairwise matched by the
edges of S, except for the root r. Hence, #(A) = #(B) +1 > #(B) (
).
5.9 Maximum Flow in a Transport Network: The FordFul-
kerson Algorithm
A transport network is a directed graph G = (V, E) with weighted arcs that satises the fol-
lowing:
(1) G is connected and loopless,
(2) G has only one source s,
(3) G has only one sink t, and
(4) the weight c(e) of the arc e is called the capacity and it is a nonnegative real number, i.e.
we have a mapping c : E R
0
.
(Compare to stationary linear networks in Section 4.4.) Actually, we could assume that G has
every possible arc except loops and it can even have multiple parallel arcs. If this is not the case,
then we simply add the missing arcs with capacity zero. Naturally, we can also assume that G
is nontrivial.
A ow f of a transport network is a weight mapping E R
0
, which satises:
(i) For each arc e, we have the capacity constraint f(e) c(e), and
(ii) each vertex v ,= s, t satises the conservation condition (also called Kirchhoff s Flow
Law, compare to Section 4.4)
initial vertex
of e is v
f(e) =
terminal vertex
of e is v
f(e).
f(e) is called the ow of e. The ow of the arc (u, v) is also denoted as f(u, v). The value of
the ow f is
[f[ =
initial vertex
of e is s
f(e).
A ow f
is a maximum ow if its value is the largest possible, i.e. [f
[ [f[ for every other

ow f.
An st cut of a transport network S is a (directed) cut I = V
1
, V
2
such that s is in V
1
and
t is in V
2
. The capacity of such a cut is
c(I) =
uV
1
vV
2
c(u, v).
(Note that the arcs in the direction opposite to the cut do not affect the capacity.) The capacity of
the cut V
1
, V
2
is also denoted as c(V
1
, V
2
). Furthermore, we dene the ux of the cut I = V
1
,
V
2
as
f
+
(I) =
uV
1
vV
2
f(u, v)
and the counter-ux as
f
(I) =
uV
2
vV
1
f(u, v).
The value of a ow can now be obtained from the uxes of any st cut:
Theorem 5.5. If f is a ow of a transport network and I is an st cut, then
[f[ = f
+
(I) f
(I).
Proof. Obviously,
initial vertex
of e is v
f(e)
terminal vertex
of e is v
f(e) =
[f[ if v = s
0 if v ,= s, t.
We denote I = V
1
, V
2
. By going through the vertices v in V
1
and by adding up the equations
we get
vV
1
initial vertex
of e is v
f(e)
vV
1
terminal vertex
of e is v
f(e) = [f[.
For each arc e whose end vertices are both in V
1
, f(e) and f(e) are added exactly once and
thus they cancel out. Therefore,
uV
1
vV
2
f(u, v)
uV
2
vV
1
f(u, v) = [f[.
Corollary. If f is a ow of a transport network and I is an st cut, then [f[ c(I).
Proof. [f[ = f
+
(I) f
(I) f
+
(I) c(I).
An arc e of a transport network is saturated if f(e) = c(e). Otherwise, it is unsaturated.
Now, we point out that [f[ = c(V
1
, V
2
) if and only if
(i) the arc (u, v) is saturated whenever u V
1
and v V
2
, and
(ii) f(u, v) = 0 whenever u V
2
and v V
1
.
An st cut I
of a transport network is called a minimum cut if c(I
) c(I) for every other

st cut I.
Corollary. If f is a ow of a transport network, I is an st cut and [f[ = c(I), then f is a
maximum ow and I is a minimum cut.
Proof. If f
is a maximum ow and I
is a minimum cut, then [f
[ c(I
) by the corollary
above. Thus,
[f[ [f
[ c(I
) c(I)
and f is indeed a maximum ow and I is indeed a minimum cut.
Actually, the value of the maximum ow is the capacity of the minimum cut. The show this, we
examine a path from vertex s to vertex v (not necessarily a directed path):
s = v
0
, e
1
, v
1
, e
2
, . . . , e
k
, v
k
= v (path p).
If e
i
= (v
i1
, v
i
), then the arc e
i
is a forward arc. If e
i
= (v
i
, v
i1
), then the arc e
i
is a back arc.
The arc e
i
of p is now weighted by the following formula:
(e
i
) =
c(e
i
) f(e
i
) if e
i
is a forward arc
f(e
i
) if e
i
is a back arc
and the path p is weighted by the following formula:
(p) =
k
min
i=1
(e
i
).
The path p is unsaturated if (p) > 0, i.e. all of the forward arcs of p are unsaturated and
f(e
i
) > 0 for all the back arcs e
i
of p.
In particular, an st path can be unsaturated in which case it is called an augmenting path
3
.
All of these denitions are of course attached to a certain ow f. By starting from an st path
p (and a ow f), we can dene a new ow:
f =
f(e) +(p) if e is a forward arc of p

f(e) (p) if e is a back arc of p
f(e) otherwise.
f is really a ow. Changes in f can only occur at the arcs and vertices of p. Every arc of
p satises the capacity constraint because of how (p) and f(e) are dened. A vertex v
i
of p
satises the conservation condition which we can verify. We have four cases:
e
i
e
i+1
v
i
e
i
e
i+1
v
i
e
i
e
i+1
v
i
e
i
e
i+1
v
i
Obviously (think about the source s)
[f[ = [f[ +(p),
so f is not a maximum ow if it has an augmenting path. Moreover, the converse is true as well.
Hence,
Theorem 5.6. A ow is a maximum ow if and only if it does not have any augmenting path.
Proof. As we claimed, a maximum ow can not have an augmenting path. Let us assume that
a ow f does not have an augmenting path. We denote the set of vertices which we can reach
from the source s along unsaturated paths by V
1
. Then, trivially, s V
1
and t / V
1
(because
there are no augmenting paths). Thus, the cut I = V
1
, V
2
is an st cut. We proceed to prove
that [f[ = c(I). By the previous corollary, f is then a maximum ow.
Let us consider the arc (u, v), where u V
1
and v V
2
. Then, there exists an unsaturated
su path p. The edge (u, v) is saturated because there would be an unsaturated sv path oth-
erwise. Similarly, we conclude that f(u, v) = 0 for every arc (u, v), where u V
2
and
v V
1
. Therefore, the ux f
+
(I) is c(I) and the counter-ux f
(I) is zero. By Theorem

5.5, [f[ = c(I).
3
Not to be confused with the augmenting path in the previous section!
We have also proven the celebrated
Theorem 5.7. (Max-Flow Min-Cut Theorem) The value of a maximum ow in a transport
network is the same as the capacity of a minimum cut.
If the capacities of the arcs are rational numbers, then a maximum ow can be found by
using Theorem 5.6. The algorithm tries to nd an augmenting path for f. If it can not be found,
then we have a maximum ow. If we nd an augmenting path, then we use it to create a greater
ow f. In the algorithm, we use a label for the vertices in the following way:
(v) = (u, direction, ),
where u is a vertex in the transport network (or if it is not dened), direction is either
forward () or back () (or if it is not dened) and is a nonnegative real number (or
). The point is, whenever a vertex v is labeled, there is an sv path p which contains the
(directed) arc (u, v) and = (p). A direction is forward if an arc is in the direction of the
path and back otherwise. We can label a vertex v when the vertex u has been labeled and either
(u, v) or (v, u) is an arc. We have two cases:
(1) (Forward Label) If e = (u, v) is an arc and (u) = (, ,
u
) as well as c(e) > f(e), then
we can write (v) = (u, ,
v
), where
v
= min
u
, c(e) f(e).
(2) (Back Label) If e = (v, u) is an arc and (u) = (, ,
u
) as well as f(e) > 0, then we
can write (v) = (u, ,
v
), where
v
= min
u
, f(e).
There are two phases in the algorithm. In the rst phase, we label the vertices as presented
above and each vertex is labeled at most once. The phase ends when the sink t gets labeled as
(t) = (, ,
t
), or when we can not label any more vertices. In the second case, there are no
augmenting paths and the ow we obtain is a maximum ow so we stop. In the rst case, the
ow we obtain is not a maximum ow and we have an augmenting path p for which (p) =
t
.
The algorithm moves on to the second phase. In the second phase, we construct a new greater
ow f by using the labels of the vertices of p obtained previously. After this, we go back to the
rst phase with this greater new ow.
The FordFulkerson Algorithm:
1. Choose an initial owf
0
. If we do not have a specic owin mind, we may use f
0
(e) = 0.
Label the source s by (s) (, , ). Set f f
0
.
2. If we have a unlabeled vertex v, which can be labeled either forward by (w, ,
v
) or
backward by (w, ,
v
), then we choose one such vertex and label it. (There can be
many ways of doing this and all of them are permitted.) If such a vertex v does not exist,
output the maximum ow f and stop.
3. If t has not been labeled, go to step #2. Otherwise, set u t.
4. If (u) = (w, ,
u
), then set
f(w, u) f(w, u) +
t
and u w.
If (u) = (w, ,
u
), then set
f(u, w) f(u, w)
t
and u w.
5. If u = s, then remove all the labels but not the label of the source and go to step #2. If
u ,= s, then go to step #4.
If f
0
(e) and c(e) are rational numbers, then the algorithm stops and produces a maximum ow.
4
In this case, we can assume that these weights and capacities are nonnegative integers. Thus,
the value of a ow increases by a positive integer every time we move from the second phase to
the rst phase and the value reaches a maximum eventually. On the other hand, the number of
steps can be as large as the value of the maximum ow. The performance time of the algorithm
does not only depend on the number of vertices but also the capacities.
The algorithm can be modied
5
so that it does not depend on the capacities. Thus, it will
work for irrational capacities. In this case, our purpose during the labeling phase is to nd the
shortest augmenting path. We get this by always choosing the vertex v in step #2 in such a way
that in (v) = (w, ,
v
), w received its label as early as possible.
The FordFulkerson Algorithm also works for nding a maximum matching in a bipartite
graph. Let us do an example:
Example. Using the bipartite graph G from an example in the previous section, we get a trans-
port network G
:
v
1
v
2
v
3
v
4
v
5
w
1
w
2
w
3
w
4
w
6
w
5
G:
G: s
t
Every edge of G
is directed from left to right and given a capacity of 1. The initial ow is a

zero ow (or a greater ow we obtain from some other initial ow). During the whole process,
the ows of the edges are integers 0 or 1. We take into the matching those edges in G whose
corresponding edges e in G
receive a ow f(e) = 1 and a maximum ow gives a maximum

matching. Note that an augmenting path can be of length larger than three in this case. (We
can also claim now that the augmenting paths here and the augmenting paths obtained from the
Hungarian Algorithm do have something in common after all!)
4
If there are irrational capacities or ows f
0
(e), then the algorithm may not stop at all and it may not produce a
maximum ow even if the process repeats endlessly. Of course, we do not have to use irrational ows. In practice,
we will not use irrational capacities.
5
This is known as the EdmondsKarp Modication (refer e.g. to SWAMY & THULASIRAMAN).
Chapter 6
Drawing Graphs
6.1 Planarity and Planar Embedding
We have not treated graphs as geometric objects so far in the course. In practice, we draw
graphs, i.e. we treat vertices as geometric points and edges as continuous curves. If a graph
G can be drawn on a plane (or a sphere) so that the edges only intersect at vertices, then it is
planar. Such a drawing of a planar graph is a planar embedding of the graph.
A connected part of a plane which does not contain any vertices and is surrounded by edges
is called a region of a planar embedding. In addition, the part outside the embedding is consid-
ered as a region, known as the exterior region (when we draw a planar graph on a plane or on a
sphere, it is just like any other region). The vertices surrounding a region s are called boundary
vertices and the edges surrounding s are called boundary edges. Two regions are adjacent if
they share a boundary edge. Note that a region can be adjacent to itself.
Example. In the following planar embedding
v
2
v
3
v
1
v
4
v
5
e
1
e
2
e
3
e
4
e
5
v
6
e
10
e
8
s
1
s
2
s
4
s
3
s
5
e
9
v
7
v
8
e
6
e
7
the regions are s
1
, s
2
, s
3
, s
4
and s
5
(the exterior region) and their boundary vertices and edges
as well as their adjacent regions are given in the table below:
region boundary vertices boundary edges adjacent regions
s
1
v
1
, v
5
, v
2
e
1
, e
10
, e
2
s
2
, s
5
s
2
v
2
, v
5
, v
4
, v
3
, v
6
, v
7
e
2
, e
4
, e
7
, e
9
, e
8
, e
6
s
1
, s
2
, s
3
, s
5
s
3
v
4
, v
5
e
4
, e
5
s
2
, s
5
s
4
v
5
e
3
s
5
s
5
v
1
, v
5
, v
4
, v
3
, v
2
, v
8
e
10
, e
3
, e
5
, e
7
, e
6
, e
1
s
1
, s
2
, s
3
, s
4
85
CHAPTER 6. DRAWING GRAPHS 86
In the following, we investigate some fundamental properties of planar embeddings of
graphs.
Theorem 6.1. (Eulers Polyhedron Formula
1
) If a planar embedding of a connected graph G
has n vertices, m edges and f regions, then
f + n = m + 2.
Proof. Let us use induction on m.
Induction Basis: m = 0. Planar embedding of G has only one vertex and one region (the
exterior region) so the claim is true.
Induction Hypothesis: The theorem is true for m . ( 0)
Induction Statement: The theorem is true for m = + 1.
Induction Statement Proof: We choose an edge e of G and examine the graph G
= G e.
If e is in a circuit, then G
is connected and by the Induction Hypothesis, we get

f
+ n = (m1) + 2,
where f
is the number of regions in G
. However, closing the circuit with e increases the

number of regions by one so f
= f 1 and the theorem is true. If G e is disconnected,

then it has two planar components, G
1
and G
2
whose number of vertices, edges and regions are
n
1
, n
2
, m
1
, m
2
, f
1
and f
2
, respectively. By the Induction Hypothesis,
f
1
+ n
1
= m
1
+ 2 and f
2
+ n
2
= m
2
+ 2.
While adding e, the number of regions becomes f
1
+f
2
1 (G
1
and G
2
share the same exterior
region or one exterior region is drawn to be a region of the other component), the number of
vertices becomes n
1
+ n
2
and the number of edges becomes m
1
+m
2
+ 1. Hence, the claim is
true.
Example. (Continuing from the previous example) We remove the vertex v
8
to get a connected
planar embedding. Now, we have 7 vertices, 10 edges, 5 regions and 5 + 7 = 10 + 2.
Theorem 6.2. (The Linear Bound) If a simple connected planar graph G has n 3 vertices
and m edges, then
m 3n 6.
Proof. If the regions of a planar embedding of G are s
1
, . . . , s
f
, then we denote the number of
boundary edges of s
i
by r
i
(i = 1, . . . , f). The case f = 1 is obvious because G is then a tree
and m = n 1 3n 6. Thus, we assume that f 2. Since G is simple, every region has at
least 3 boundary edges and thus
f
i=1
r
i
3f.
Every edge is a boundary edge of one or two regions in the planar embedding, so
f
i=1
r
i
2m.
The result now follows directly from Eulers Polyhedron Formula.
1
The name comes from a polyhedron with n vertices, m edges, f faces and no holes.
Theorem 6.3. (The Minimum Degree Bound) For a simple planar graph G, (G) 5.
Proof. Let us prove by contradiction and consider the counter hypothesis: G is a simple planar
graph and (G) 6. Then, (by Theorem 1.1) m 3n, where n is the number of vertices and
m is the number of edges in G. (
Theorem 6.2)
A characterization of planar graphs is obtained by examining certain forbidden subgraphs.
Theorem 6.4. (Kuratowskis Theorem) A graph is planar if and only if none of its subgraphs
can be transformed to K
5
or K
3,3
by contracting edges.
The proof is quite complicated (but elegant!), refer e.g. to SWAMY & THULASIRAMAN for
more information. K
5
and K
3,3
are not planar, which can be veried easily.
There are many fast but complicated algorithms for testing planarity and drawing planar
embeddings. For example, the HopcroftTarjan Algorithm
2
is one. We present a slower classi-
cal polynomial time algorithm, the DemoucronMalgrangePertuiset Algorithm
3
(usually just
called Demoucrons Algorithm). The idea of the algorithm is to try to draw a graph on a plane
piece by piece. If this fails, then the graph is not planar.
If G is a graph and R is a planar embedding of a planar subgraph S of G, then an R-piece
P of G is
either an edge of G S whose end vertices are in S, or
a component of the subgraph induced by vertices not in S which contains the edges (if
any) that connect S to the component, known as pending edges, and their end vertices.
Those vertices of an R-piece of G that are end vertices of pending edges connecting them to
S are called contact vertices. We say that a planar embedding R of the planar subgraph S is
planar extendable to G if R can be extended to a planar embedding of the whole G by drawing
more vertices and/or edges. Such an extended embedding is called a planar extension of R to
G. We say further that an R-piece P of G is drawable in a region s of R if there is a planar
extension of R to G where P is inside s. Obviously all contact vertices of P must then be
boundary vertices of s, but this is of course not sufcient to guarantee planar extendability of R
to G. Therefore we say that a P is potentially drawable in s if its contact vertices are boundary
vertices of s. In particular, a piece with no contact vertices is potentially drawable in any region
of R.
Demoucrons Algorithm:
1. We rst check whether or not G is a forest. If it is a forest, then it clearly is planar and
can be planar embedded. (There are fast algorithms for this purpose.) We can then stop.
2. If G is not a forest then it must contain at least one circuit. We choose a circuit C, embed
it to get the planar embedding D, and set R D. (A circuit is obviously planar and is
easily planar embedded.)
3. If R is a planar embedding of G, then we output it and stop.
2
The original reference is HOPCROFT, J.E. & TARJAN, R.E.: Efcient Planarity Testing. Journal of the ACM
21 (1974), 549568.
3
The original reference is DEMOUCRON, G. & MALGRANGE, Y. & PERTUISET, R.: Graphes planaires: recon-
naissance et construction des reprsentations planaires topologiques. Revue Franaise Recherche Oprationnelle
8 (1964), 3347.
4. We construct the set T of all R-pieces of G. For each piece P T we denote by o(P)
the set of all those regions of R which P is potentially drawable in.
5. If, for an R-piece P T, the set o(P) is empty then G is not planar. We can then output
this information and stop.
6. Choose an R-piece P, starting from those potentially drawable only in one region.
7. Depending on the number of contact vertices of P, we planar extend R:
7.1 If P has no contact vertices, we call Demoucrons Algorithm recursively with input
P. If it turns out that P is not planar, then G is not planar, and we output this
information and stop. Otherwise we extend R to a planar embedding U by drawing
P in one of its regions, set R U, and return to step #3.
7.2 If P has exactly one contact vertex v, with the corresponding pendant edge e, we
call Demoucrons Algorithm recursively with input P. If it turns out that P is not
planar, then G is not planar, and we output this information and stop. Otherwise we
extend Rto a planar embedding U by drawing P in a region with boundary vertex v,
set R U, and return to step #3. (This region of R will then be an exterior region
of the planar embedding of P.)
7.3 If P has (at least) two contact vertices v
1
and v
2
, they are connected by a path p in
P. We then extend R to a planar embedding U by drawing p in a region of R with
boundary vertices v
1
and v
2
where P is potentially drawable, set R U, and return
to step #3.
Clearly, if G is not planar, Demoucrons Algorithm will output this information. On the
other hand, the algorithm will not get stuck without drawing the planar embedding if the input
is planar, because
Statement. If G is planar, then at each step of the algorithm R is planar extendable to G.
Proof. We use induction on the number of times the algorithm visits step #7.
Induction Basis: = 0. Now either G is a forest (and R is not needed) or R is a circuit of
G. Obviously the planar embedding of this circuit can be planar extended to G.
Induction Hypothesis: The statement is true for r. (r 0)
Induction Statement: The statement is true for = r + 1.
Incuction Statement Proof: For step #7.1 the matter is clear. If, in step #7.2, P is potentially
drawable in the region s of R, it can always be drawn in this region without endangering subse-
quent steps. In other words, any possible region can be chosen. This is because the region can
be exchanged for another at all times by reection with respect to the vertex v (and possibly
rescaling):
v
e
P
s
1
s
2
s
3
v
e
P
s
1
s
2
s
3
Similarly, if in step #7.3, P is drawable in a region of R, then it can be drawn in this region
without endangering subsequent steps. If P is drawable in both region s
1
and region s
2
, its
contact vertices are boundary vertices of both s
1
and s
2
. At any time, a drawn P (or part of
it) can be moved from region s
1
to s
2
, or vice versa, simply by reection with respect to the
common boundary (and possibly rescaling to t into the region).
Remark. Nonplanar graphs may be embedded on closed continuous surfaces with holes. For
instance, a torus is closed surface with exactly one hole. On a torus we can embed the non-
planar graphs K
5
, K
6
and K
7
, and also K
3,3
. K
8
is more complex and its embedding requires
a closed surface with two holes. The smallest number of holes in a closed surface required for
embedding the graph G on it is called the genus of G. On the other hand the smallest number of
crossings of edges in a drawing of G on plane is called the crossing number of G. Computation
of genus and crossing number are both AT-complete problems.
A coloring of a graph is a labeling of vertices where adjacent vertices never share a label.
The labels are then often called colors. We say that a graph is k-colorable if it can be colored
using (at most) k colors. If a graph is colorable then it obviously can not have loops. Equally
obviously, parallel edges can be reduced to one, so we may assume our graphs here to be simple.
The smallest number k for which the graph G is k-colorable, is called the chromatic number of
G, denoted by (G).
K
4
is an example of a planar simple graph which is not 3-colorable. On the other hand there
is the celebrated
Theorem 6.5. (The Four-Color Theorem) Every simple planar graph is 4-colorable.
Proof. The only known proofs require extensive computer runs. The rst such proof was ob-
tained by Kenneth Appel ja Wolfgang Haken in 1976. It takes a whole book to present the
proof: APPEL, K. & HAKEN, W.: Every Planar Map is Four Colorable. American Mathemat-
ical Society (1989).
If we require a bit less, i.e. 5-colorability, then there is much more easily provable result,
and an algorithm.
Theorem 6.6. (Heawoods Theorem or The Five-Color Theorem) Every simple planar graph
is 5-colorable.
Proof. We may think of Gas a planar embedding. We use induction on the number n of vertices
of G.
Induction Basis: n = 1. Our graph is now 1-colorable since there are no edges.
Induction Hypothesis: The theorem is true for n . ( 1)
Induction Statement: The theorem is true for n = + 1.
Induction Statement Proof: According to the Minimum Degree Bound, there is a vertex v
in G of degree at most 5. On the other hand, according to the Induction Hypothesis the graph
G v is 5-colorable. If, in this coloring, the vertices adjacent to v are colored using at most
four colors, then clearly we can 5-color G.
So we are left with the case where the vertices v
1
, v
2
, v
3
, v
4
, v
5
adjacent to v are colored
using different colors. We may assume that the indexing of the vertices proceeds clockwise,
and we label the colors with the numbers 1, 2, 3, 4, 5 (in this order). We show that the coloring
of Gv can be changed so that (at most) four colors sufce for coloring v
1
, v
2
, v
3
, v
4
, v
5
.
We denote by H
i,j
the subgraph of G v induced by the vertices colored with i and j. We
have two cases:
v
1
and v
3
are in different components H
1
and H
3
of H
1,3
. We then interchange the colors 1
and 3 in the vertices of H
3
leaving the other colors untouched. In the resulting 5-coloring
of Gv the vertices v
1
and v
3
both have the color 1. We can then give the color 3 to v.
v
1
and v
3
are connected in H
1,3
. Then there is a v
1
v
3
path in H
1,3
. Including the vertex
v we get from this path a circuit C. Now, since we indexed the vertices v
1
, v
2
, v
3
, v
4
, v
5
clockwise, exactly one of the vertices v
2
and v
4
is inside C. We deduce that v
2
and v
4
are
in different components of H
2,4
, and we have a case similar to the previous one.
The proof gives a simple (recursive) algorithm for 5-coloring a planar graph, the so-called
Heawoods Algorithm.
6.2 The DavidsonHarel Algorithm
For the actual drawing of a graph we need to dene the drawing area (the window), i.e. a
rectangular area with sides parallel to the coordinate axes, the drawing curve of the edges (here
edges are drawn as line segments), and certain criteria of beauty, so that the resulting drawing
is pleasant to the eye, balanced, and as clear as possible. Such beauty criteria are of course
context-dependent and even matters of individual taste. In the sequel we restrict ourselves to
simple graphs, given by, say, an adjacency matrix or an all-vertex incidence matrix.
We will now present the so-called DavidsonHarel Algorithm
4
which, applying an anneal-
ing algorithm, aims at better and better drawings of a graph using a certain ugliness function
(cf. Section 5.7). An ugliness function R computes a numerical ugliness value obtained from
a drawing P of a graph G. This value is a sum of various contributing factors. We denote, as
usual, the sets of vertices and edges of G by v
1
, . . . , v
n
and e
1
, . . . , e
m
, respectively. We
also denote by v
i
the vector (or geometric point) corresponding to the vertex v
i
, and by e
j
the
line segment corresponding to the edge e
j
. Further, we denote
d
ij
= |v
i
v
j
|,
r
i
= distance of v
i
from the right border of the window,
l
i
= distance of v
i
from the left border of the window,
u
i
= distance of v
i
from the upper border of the window,
b
i
= distance of v
i
from the lower border of the window,
c
j
= length of the line segment e
j
,
f
ij
=
_
1, if the line segments e
i
and e
j
intersect without e
i
and e
j
being adjacent
0 otherwise,
g
ij
=
_
distance of v
i
from the line segment e
j
if it exceeds and v
i
is not an end vertex of e
j
otherwise.
is a parameter of the algorithm telling how close to vertices edges can be. The ugliness
function is then given by
4
The original reference is DAVIDSON, R. & HAREL, D.: Drawing Graphs Nicely Using Simulated Annealing.
ACM Transactions on Graphics 15 (1996), 301331.
R(P) =
1
n1
i=1
n
j=i+1
1
d
2
ij
+
2
n
i=1
_
1
r
2
i
+
1
l
2
i
+
1
u
2
i
+
1
b
2
i
_
+
3
m
j=1
c
2
j
+
4
m1
i=1
m
j=i+1
f
ij
+
5
n
i=1
m
j=1
1
g
2
ij
,
where
1
, . . . ,
5
are nonnegative-valued parameters weighting the contributions of the various
factors. (One could actually use negative values as well, whatever the interpretation then might
be.)
We can compute d
ij
, . . . , g
ij
quite easily using some basic formulae of vector geometry. We
must, however, think about the speed of the computation as well. One way to speed up the
computation is to use complex arithmetic. d
ij
, . . . , g
ij
are then equally easily computable.
5
It
may also be of advantage to force the vertices into a lattice of geometric points. This can be
achieved for instance by rounding the coordinates (or complex numbers) to a xed accuracy
and abandoning drawings where the ugliness function has the value (this happens e.g. when
vertices occupy the same point).
In the annealing process the state is P and the response is R(P). An initial state can be
obtained by choosing the points v
1
, . . . , v
n
in the window randomly, and then drawing the
edges accordingly. The state transition process P A
(P) is the following:

Choose a random vertex v
i
. (Alternatively the vertices may be circulated cyclically.)
Draw a circle of radius centered on v
i
. The radius is a parameter, which is initially
large and gradually reduced later in some systematic fashion.
Chose a random point u on this circle.
If u is outside the drawing window, the state remains the same. Otherwise set v
i
u
and change the edges accordingly in the drawing.
The remaing parts of the algorithm are very similar to the annealing algorithm for the TSP in
Section 5.7.
Remark. This method has numerous variants. The window could be a circle and the edges
concentric arcs or radii. Or the window could be a sphere and edges drawn as arcs of great
circles. The window could also be unbounded, for instance, the whole of R
2
. We could draw
graphs three-dimensionally. Etc. We could also use a metric other than the Euclidean one when
computing distances, e.g. the Manhattan metric (1-norm) or the max-metric (-norm),
geodetic distances on a sphere, etc. Needless to say, the resulting drawings are rather different
using these variants of the algorithm.
It may be noted that using nearly any effective criteria, nding the optimally pleasing draw-
ing of a simple graph is an AT-hard problem.
5
Note that if z
1
= x
1
+jy
1
and z
2
= x
2
+jy
2
, where j is the imaginary unit, then the real part of z
1
z
2
equals
the dot product (x
1
, y
1
) (x
2
, y
2
) and the imaginary part equals the determinant
x
1
x
2
y
1
y
2
.
Chapter 7
MATROIDS
Many concepts in the preceding chapters do not so much deal with graphs themselves as their
structural properties. Examples are various dualities (cut set vs. circuit), principles behind cer-
tain algorithms (e.g. Kruskals Algorithms), and various extremality properties (many structures
are the smallest of their kind, one cannot e.g. remove an edge of a cut set without it losing
this property).
Exactly corresponding structures were found in many other areas of mathematics, and they
were called matroids.
1
7.1 Hereditary Systems
A hereditary family of sets is a family of sets such that whenever a set F is in the family then
so are all subsets of F (and in particular the empty set ). A hereditary system M of a set E
is a nonempty hereditary family of subsets of E. Included there are also the various ways of
specifying J
M
, called aspects. It will be assumed in what follows that E is a nite set. The
following nomenclature is traditional:
Sets in the family J
M
are called independent sets of M.
The family of subsets of E other than those in J
M
is denoted by T
M
and called the family
of dependent sets of M.
An independent set is maximal if it is not a proper subset of another independent set. A
maximal independent set is called a basis. The family of all bases is denoted by B
M
. Note
that an independent set is always contained in a basis.
A dependent set is minimal if no dependent set is its proper subset. A minimal dependent
set is called a circuit.
2
(Recall that the empty set is always in J
M
.) The family of all
circuits is denoted by (
M
. Note that a dependent set always contains a circuit.
A circuit consisting of only one element is a so-called loop. Elements of a circuit with
two elements are called parallel. A hereditary system is simple if it has no loops and no
parallel elements.
1
The remarkable thing is that many of these structures were found independently at the same time around the
year 1935: Hassler Whitney investigated planarity of graphs, Saunders MacLane geometric lattices of points, and
Bartel van der Waerdens topic was independence in vector spaces.
2
This or any other familiar sounding concept should not be confused with the corresponding concept for
graphs, even though there is a certain connection, as will be seen!
92
CHAPTER 7. MATROIDS 93
The rank of a subset F of E is the largest size of an independent set contained in F.
(Recall that E is assumed to be nite.) Note that the empty set is always an independent
set contained in F. The rank of F is denoted by
M
(F), and
M
is called the rank function
of M.
A notation similar to one used for graphs will be adopted in the sequel concerning adding
an element e to the set F (denoted by F + e) or removing it from F (denoted by F e). Two
easy properties of the rank function are the following
Theorem 7.1. If M is a hereditary system of the set E then
(i)
M
() = 0, and
(ii) for any subset F of E and any element e,
M
(F)
M
(F + e)
M
(F) + 1.
Proof. Item (i) is clear, so let us move to item (ii).
Since F +e contains those independent sets that are contained in F, we have
M
(F +e)
M
(F). On the other hand, possible independent subsets of F +e not contained in F may only
consist of an independent subset of F and e, so
M
(F + e)
M
(F) + 1.
A hereditary system M may of course be specied by giving its independent sets, that is by
giving J
M
. It can be specied as well by giving its bases, i.e. B
M
, independent sets will then
be exactly all subsets of bases. On the other hand, M can be specied by giving its circuits,
i.e. (
M
, independent sets are then the sets not containg circuits. Finally, M can be dened by
giving the rank function
M
, since a set F is independent exactly when
M
(F) = #(F). (As
before, we denote cardinality of a set F by #(F).) Thus an aspect may involve any of J
M
, B
M
,
(
M
and
M
.
It might be mentioned that a hereditary system is a far too general concept to be of much
use. This means that well chosen aspects are needed to restrict the concept to a more useful one
(that is, a matroid). Let us have a look at certain proper aspects in connection with a matroid
well familiar from the preceding chapters.
7.2 The Circuit Matroid of a Graph
The circuit matroid M(G) of a graph G = (V, E) is a hereditary system of the edge set E
whose circuits are the circuits of G, considered as edge sets. (It is naturally assumed that G is
not empty.) The bases of M(G) are the maximal independent edge sets, i.e. spanning forests
of G, and the independent sets of M(G) are the subforests, both considered as edge sets. Let
us denote G
F
= (V, F) for a subset F of E. The number of vertices of G is denoted by n, as
usual.
Remark. A hereditary system that is not directly a circuit matroid of any graph but has a
structure identical to one is called a graphic matroid.
Let us then take a look at different aspects of the circuit matroid.
Basis Exchange Property
Let us consider two bases (i.e. spanning forests) B
1
and B
2
. If e is an edge in B
1
, its removal
divides some component G
of the graph G into two disjoint subgraphs. Now certain edges of

B
1
will be the branches of a spanning tree T
1
of G
, and similarly, certain edges in B

2
will be
the branches of a spanning tree T
2
of G
. The removed edge e is either a branch of T

2
or then a
link of T
2
. In the latter case e will be in the fundamental cut set determined by a branch f of T
2
(cf. Theorem 2.7). Then T
1
e +f is also a spanning tree of G
and we can replace e by f and

get again a spanning forest of G, that is, a basis.
Hence we have
Basis Exchange Property: If B
1
and B
2
are different bases and e B
1
B
2
then there is an
element f B
2
B
1
such that B
1
e + f is a basis.
In general, a hereditary system with the basis exchange property will be a matroid. In other
words, the basis exchange property is a proper aspect. Using basis exchange one can move
from one basis to another. All bases are thus of the same size.
Uniformity. Absorptivity
For a subset F of E let us denote by n
F
the number of vertices in the subgraph F of G
induced by F, and by k
F
the number of its components. Then there are n
F
k
F
edges in a
spanning forest of F. Let us denote further by K
F
the number of components of the subgraph
G
F
of G. Clearly then
M(G)
(F) = n
F
k
F
= n K
F
,
and all such spanning forests are of the same size. Hence
Uniformity: For a subset F of E all maximal independent subsets of F are of the same size.
(Maximality of a set means here that there are no independent sets J such that H J F.)
In general, a hereditary system with the uniformity property will be a matroid, and uniformity
is a proper aspect.
In the gure below continuous lines are the edges of F, with the thick ones being the
branches of a spanning forest. Dashed lines indicate the remaining edges in E.
e
f
If e is an edge of G and
M(G)
(F + e) =
M(G)
(F) then e does not connect two components
of G
F
. Suppose f is another edge with the same property, that is,
M(G)
(F + f) =
M(G)
(F).
Clearly then
M(G)
(F + e + f) =
M(G)
(F).
Thus we get
Weak Absorptivity: If e, f E and F E and
M
(F) =
M
(F + e) =
M
(F + f)
then also
M
(F + e + f) =
M
(F).
In general a weakly absorptive hereditary system is a matroid, and thus weak absorptivity is
another proper aspect.
By repeating the above argument sufciently many times we see that if F and F
are sets of
edges of G, and for each edge e in F
we have
M(G)
(F + e) =
M(G)
(F), then
M(G)
(F F
) =
M(G)
(F).
Hence also
Strong Absorptivity: If F, F
E and
M
(F + e) =
M
(F) for each element e in F
then
M
(F F
) =
M
(F).
We conclude that strong absorptivity is a proper aspect.
Augmentation
Suppose I
1
and I
2
are independent sets of the circuit matroid M(G) (edge sets of subforests
of G) and #(I
1
) < #(I
2
). The subgraph G
I
1
then has n#(I
1
) components, and the subgraph
G
I
2
has n #(I
2
) components, so strictly less than G
I
1
. Adding an edge does not reduce the
number of components exactly in the case where the edge is added in some component. Thus,
if adding any edge in I
2
I
1
to G
I
1
preserves the number of components then it must be that
the edge is added in some component of G
I
1
, and G
I
2
cannot have fewer components than G
I
1
.
But as noted, this is not the case if #(I
1
) < #(I
2
), and so
Augmentation: If I
1
and I
2
are independent sets of the hereditary system M and #(I
1
) <
#(I
2
) then there exists an element e I
2
I
1
such that I
1
+ e is in J
M
.
In general, a hereditary system with the augmentation property is a matroid. Thus augmentation
is a proper aspect, too.
Elimination
The circuits of the circuit matroid M(G) are the edge sets of the circuits of G. The degree
of a vertex in a circuit is two. If C
1
and C
2
are different circuits of M(G) then the degree of
a vertex of the ring sum C
1
C
2
is also even, see Section 1.3. Hence C
1
C
2
must
contain at least one circuit as a subgraph, since a ring sum does not have isolated vertices and
a nonempty forest has at least one pending vertex (Theorem 2.3). Recalling the denition of
ring sum in Section 1.3 it is noticed that such a circuit does not contain edges in the intersection
C
1
C
2
, at least not with as high multiplicity as in C
1
C
2
. Thus
Elimination Property: If C
1
and C
2
are different circuits of the hereditary system M and
e C
1
C
2
then there is a circuit C (
M
such that C C
1
C
2
e.
Again, elimination property is a proper aspect, and a hereditary system with the elimination
property is a matroid.
Induced Circuits
If I is an independent set of the circuit matroid M(G) (edge set of a subforest) then adding
one edge either closes exactly one circuit in a component of G
I
(Theorem 2.3), or then it con-
nects two components of G
I
and does not create a circuit. We have then
Property of Induced Circuits: If I is an independent set of a hereditary system M and e E
then I + e contains at most one circuit.
The property of induced circuits is a proper aspect, and a hereditary system having this property
will be a matroid.
7.3 Other Basic Matroids
Vectorial Matroid
Let E be a nite set of vectors of a vector space (say R
n
) and the independent sets of a
hereditary systemM of E be exactly all linearly independent subsets of E (including the empty
set). M is then a so-called vectorial matroid. Here E is usually allowed to be a multiset, i.e.
its elements have multiplicitiescf. parallel edges of graphs. It is then agreed that a subset of
E is linearly dependent when one its elements has a multiplicity higher than one. A hereditary
system that is not directly vectorial but is structurally identical to a vectorial matroid M
is
called a linear matroid, and the matroid M
is called its representation.

A circuit of a vectorial matroid is a linearly dependent set C of vectors such that remov-
ing any of its elements leaves a linearly independent setkeeping in mind possible multi-
ple elements. An aspect typical to vectorial matroids is the elimination property. If C
1
=
r, r
1
, . . . , r
k
and C
2
= r, r
1
, . . . , r
l
are different circuits sharing (at least) the vector r then
r can be represented as linear combinations of other vectors in both C
1
and C
2
, and in such a
way that all coefcients in the combinations are nonzero. We get thus an equality
k
i=1
c
i
r
i

l
j=1
c
j
r
j
= 0.
Combining (possible) repetitive vectors on the left hand side, and noticing that this does not
make it empty, we see that C
1
C
2
r contains a circuit. (Note especially the case where either
C
1
= r, r or C
2
= r, r.)
In the special case where E consists of columns (or rows) of a matrix A, a vectorial matroid
of E is called a matrix matroid and denoted by M(A). For example, the circuit matroid M(G)
of a graph G is a linear matroid whose representation is obtained using the rows of the circuit
matrix of G in the binary eld GF(2) (see Section 4.5).
3
Of course, if desired, any vectorial
matroid of E may be considered as a matrix matroid simply by taking the vectors of E as
columns (or rows) of a matrix.
4
3
Hereditary systems with a representation in the binary eld GF(2) are called binary matroids. The circuit
matroid of a graph is thus always binary.
4
This actually is the origin of the name matroid. A matroid is a generalization of a linear matroid and a linear
matroid may be thought of as a matrix. Indeed, not all matroids are linear. The name matroid was strongly
opposed at one time. Even today there are people who prefer to use names like geometry or combinatorial
geometry.
Transversal Matroid
Let / = A
1
, . . . , A
k
be a family of nonempty nite sets. The transversal matroid M(/)
is a hereditary system of the set E = A
1
A
k
whose independent sets are exactly all
subsets of E containing at most one element of each of the sets A
i
(including the empty set).
Here it is customary to allow the family / to be a multiset, that is, a set A
i
may appear several
times as its element, thus allowing more than one element of A
i
in an independent set.
A natural aspect of transversal matroids is augmentation, and it is connected with aug-
mentings of matchings of bipartite graphs! (See Section 5.8.) Let us dene a bipartite graph
G = (V, E) as follows: The vertex set is V = E /, and the vertices e and A
j
are connected
by an edge exactly when e A
j
. (Note how the vertex set V is naturally divided into the two
parts of the cut, E and /.) An independent set of M(/) is then a set of matched vertices of G
in E, and vice versa.
Example. In the gure below is the bipartite graph corresponding to the transversal matroid
of the family 1, 2, 2, 3, 4, 4, 5, and its independent set 1, 2, 4 (thick line).
1
2
3
4
5
{1,2}
{2,3,4}
{4,5}
Very much in the same way as in the proof of Theorem 5.3 one may show that if I
1
and I
2
are
independent sets (vertex sets of the matchings S
1
and S
2
) and #(I
1
) < #(I
2
) then there is an
augmenting path of the matching S
1
such that the new matched vertex is in I
2
. Thus M(/)
indeed has the augmentation property.
Remark. For matchings of bipartite graphs the situation is completely general. That is, match-
ings of bipartite graphs can always be thought of as independent sets of transversal matroids.
In fact this remains true for matchings of general graphs, too, leading to the so-called matching
matroids, see e.g. SWAMY & THULASIRAMAN.
If the sets of the family / are disjointi.e. they form a partition of Ethen the transversal
matroid is also called partition matroid. For a partition matroid augmentation is obvious.
Uniform Matroid
For all nite sets E one can dene the so-called uniformmatroids. The uniformmatroid of E
of rank k, denoted U
k
(E), is a hereditary system whose independent sets are exactly all subsets
of E containing at most k elements. The bases of U
k
(E) are those subsets containing exactly
k elements, and the circuits are the subsets containing exactly k + 1 elements. In particular, all
subsets of E form a uniform matroid of E of rank #(E), this is often called the free matroid of
E. Quite obviously U
k
(E) has the basis exchange property and the augmentation property.
Uniform matroids are not very interesting as such. They can be used as building blocks
of much more complicated matroids, however. It may also be noted that uniform matroids are
transversal matroids (can you see why?).
7.4 Greedy Algorithm
Many problems of combinatorial optimization
5
may be thought of as nding a heaviest or a
lightest independent set of a hereditary system M of E, when each element of E is given a
weight. The weighting function is : E R and the weight of a set F E is
eF
(e).
The two optimization modes are interchanged when the signs of the weights are reversed.
One may also nd the heaviest or the lightest bases. Again reversing the signs of the weights
interchanges maximization and minimization. If all bases are of the same sizeas will be the
case for matroidsthey can be restricted to the case where there weights are positive. Indeed,
if A is the smallest weight of an element of E then changing the weight function to
: (e) = 1 + (e) A
one gets an equivalent optimization problem with positive weights. On the other hand, maxi-
mization and minimization are interchanged when the weighting function is changed to
: (e) = 1 + B (e)
where B is the largest weight of an element of E.
Example. (A bit generalized) Kruskals Algorithm (see Section 5.6) nds a lightest spanning
forest of an edge-weighted graph G, i.e. a lightest basis of the circuit matroid of G. As was seen,
this can be done quite fastand even faster if the edges are given in the order of increasing
weight when one can always consider the best remaining edge to be included in the forest.
Kruskals Algorithm No. 1 is an example of a so-called greedy algorithm that always proceeds
in the best available direction. Such a greedy algorithm is fast, indeed, it only needs to nd
this best element to be added in the set already constructed.
It might be mentioned that Kruskals Algorithm No. 3 is also a greedy algorithm, it nds a
heaviest cospanning forest in the dual matroid of the circuit matroid, the so-called bond matroid
of G (see Section 7.6).
Even though greedy algorithms produce the correct result for circuit matroids they do not
always do so.
Example. Finding a lightest Hamiltonian circuit of an edge-weighted graph G may also be
thought of as nding the lightest basis of a hereditary systemassuming of course that there
are Hamitonian circuits. The set E is again taken to be the edge set of G but now the bases
are the Hamiltonian circuits of G (considered as edge sets). A lightest basis is then a lightest
Hamiltonian circuit. As was noticed in Section 5.7, nding a lightest Hamiltonian circuit is a
well-known AT-complete problem and no greedy algorithm can thus always produce a (cor-
rect) resultat least if T ,= AT. The hereditary system thus obtained is in general a matroid,
however (e.g. it does not generally have the basis exchange property).
It would thus appear thatat least for matroidsgreedy algorithms are favorable methods
for nding heaviest/lightest bases (or independent sets). Indeed, matroids are precisely those
hereditary systems for which this holds true. To be able to proceed further we dene the greedy
algorithm formally. We consider rst maximization of independent sets, minimization is given
in brackets. The input is a hereditary system M of the set E, and a weighting function .
5
These problems are dealt with more extensively in the course Optimization Theory 2.
Greedy Algorithm for Independent Sets:
1. Sort the elements e
1
, . . . , e
m
of E according to decreasing [increasing] weight: e
(1)
, . . . ,
e
(m)
.
2. Set F and k 1.
3. If (e
(k)
) 0 [(e
(k)
) 0], return F and quit.
4. If (e
(k)
) > 0 [(e
(k)
) < 0] and F e
(k)
is independent, set F F e
(k)
.
5. If k = m, return F and quit. Else set k k + 1 and go to #3.
For bases the algorithm is even simpler:
Greedy Algorithm for Bases:
1. Sort the elements e
1
, . . . , e
m
of E according to decreasing [increasing] weight: e
(1)
, . . . ,
e
(m)
.
2. Set F and k 1.
3. If F e
(k)
is independent, set F F e
(k)
.
4. If k = m, return F and quit. Else set k k + 1 and go to #3.
The main result that links working of greedy algorithms and matroids is
Theorem 7.2. (Matroid Greediness Theorem) The greedy algorithmproduces a correct heav-
iest independent set of a hereditary system for all weight functions if and only if the system is
a matroid. (This is the so-called greediness property.) The corresponding result holds true for
bases, and also for nding lightest independent sets and bases. Furthermore, in both cases it
sufces to consider positive weights.
Proof. The rst sentence of the theorem is proved as part of the proof of Theorem 7.3 in the
next section.
As noted above, greediness is equivalent for maximization and minimization, for both inde-
pendent sets and bases. It was also noted that nding a heaviest basis may be restricted to the
case of positive weights. Since for positive weights a heaviest independent set is automatically
a basis, greediness for bases follows from greediness for independent sets.
On the other hand, if greediness holds for bases, it holds for independent sets as well. Max-
imization of independent sets using the weight function then corresponds to maximization of
bases for the positive weight function
: (e) = 1 + max(0, (e)),
the greedy algorithms behave exactly similarly, item #3 is not activated for independent sets.
Elements of weight 1 should be removed from the output.
Remark. Greediness is thus also a proper aspect for matroids. For hereditary families of sets
it is equivalent to usefulness of the greedy algorithm. Certain other similar but more general
families of sets have their own greediness theorems. Examples are the so-called greedoids
and matroid embeddings.
7.5 The General Matroid
Any one of the several aspects above makes a hereditary system a matroid. After proving that
they are all equivalent, we may dene a matroid as a hereditary system that has (any) one of
these aspects.
Before that we add one aspect to the list, which is a bit more difcult to prove directly for
circuits matroids of graphs:
Submodularity: If M is a hereditary system of the set E and F, F
E then
M
(F F
) +
M
(F F
)
M
(F) +
M
(F
).
Let us then prove the equivalences, including submodularity.
Theorem 7.3. If a hereditary system has (any) one of the nine aspects below then it has them
all (and is a matroid).
(i) Uniformity (vi) Submodularity
(ii) Basis exchange property (vii) Elimination property
(iii) Augmentation property (viii) Property of induced circuits
(iv) Weak absorptivity (ix) Greediness
(v) Strong absorptivity
Proof. The implications are proved following the strongly connected digraph below:
(i) (ii) (iii) (iv) (v)
(vi)
(vii) (viii) (ix)
All nine aspects are then connected by implication chains in both directions, and are thus logi-
cally equivalent. Let us consider a general hereditary system M of the set E.
(i)(ii): As a consequence of uniformity, all bases of M are of the same size. If B
1
, B
2

B
M
and e B
1
B
2
, we may apply uniformity to the set F = (B
1
e) B
2
. All maximal
independent sets included in F are then of the same size as B
2
(and B
1
). Now B
1
e is not
one of these maximal sets having too few elements. On the other hand, by adding one element
f to B
1
e we get such an independent set H. The element f must then be in the set difference
B
2
B
1
, so H = B
1
e + f. Moreover, H has as many elements as B
1
, and so it is a basis.
(ii)(iii): If I
1
, I
2
J
M
and #(I
1
) < #(I
2
), we choose bases B
1
and B
2
such that I
1
B
1
and I
2
B
2
. Applying basis exchange (repeatedly) we replace those elements of B
1
I
1
that
are not in B
2
by elements of B
2
. After this operation we may assume that B
1
I
1
B
2
. As a
consequence of the basis exchange property all bases are of the same size. Thus
#(B
1
I
1
) = #(B
1
) #(I
1
) > #(B
2
) #(I
2
) = #(B
2
I
2
),
and B
1
I
1
cannot be included in B
2
I
2
. Therefore there is an element e of B
1
I
1
in I
2
and
I
1
+ e is an independent set.
(iii)(iv): Let us consider a situation where
M
(F) =
M
(F + e) =
M
(F + f).
If now
M
(F +e +f) >
M
(F), we take a maximal independent subset I
1
of F and a maximal
inpendent subset I
2
of F + e + f. Then #(I
2
) > #(I
1
) and by the augmentation property I
1
can be augmented by an element of I
2
. This element cannot be in F (why not?), so it must be
either e or f. But then
M
(F) <
M
(F + e) or
M
(F) <
M
(F + f) (
).
(iv)(v): Let us assume weak absorptivity and consider subsets F and F
of E such that
M
(F +e) =
M
(F) for each element e of F
. We use induction on k = #(F
F) and show
that
M
(F) =
M
(F F
) (strong absorptivity).
Induction Basis: Now k = 0 or k = 1 and the matter is clear.
Induction Hypothesis: The claimed result holds true when k . ( 1)
Induction Statement: The claimed result holds true when k = + 1.
Induction Statement Proof: Choose distinct elements e, f F
F and denote F
= F
e f. The Induction Hypothesis implies that
M
(F) =
M
(F F
) =
M
(F F
+ e) =
M
(F F
+ f).
Applying weak absorptivity to this it is seen that
M
(F) =
M
(F F
+ e + f) =
M
(F F
).
(v)(i): If I is a maximal independent subset of F then
M
(I+e) =
M
(I) for elements e in
the set difference F I (if any). Strong absorptivity implies then that
M
(F) =
M
(I) = #(I),
i.e. all these independent sets are of the same size and uniformity holds true.
(i)(vi): Let us consider sets F, F
E and denote by I
1
a maximal independent subset of
the intersection F F
and by I
2
a maximal inpendent subset of the union F F
. Uniformity
implies augmentation, so we may assume that I
2
is obtained from I
1
by adding elements, that
is I
1
I
2
. Now I
2
F is an independent subset of F and I
2
F
is an independent subset of
F
, and both of them include I

1
. So
M
(F F
) +
M
(F F
) = #(I
1
) + #(I
2
)
= #(I
2
F) + #(I
2
F
)
M
(F) +
M
(F
).
The equality marked by an asterisk is a set-theoretical one, see the gure below.
I
1
I
2
I
2
F F
(vi)(vii):Let us consider distinct circuits C
1
, C
2
(
M
and an element e C
1
C
2
. Then
M
(C
1
) = #(C
1
) 1 and
M
(C
2
) = #(C
2
) 1, and
M
(C
1
C
2
) = #(C
1
C
2
). (Remember
that every proper subset of a circuit is independent.) If nowC
1
C
2
e does not contain a circuit,
it is independent and
M
(C
1
C
2
e) = #(C
1
C
2
)1, whence
M
(C
1
C
2
) #(C
1
C
2
)1.
Submodularity however implies that
M
(C
1
C
2
) +
M
(C
1
C
2
)
M
(C
1
) +
M
(C
2
),
and further that (check!)
#(C
1
C
2
) + #(C
1
C
2
) #(C
1
) + #(C
2
) 1.
This is a set-theoretical impossibility, and thus C
1
C
2
e does contain a circuit.
(vii)(viii): If I is an independent set and I + e contains two distinct circuits C
1
and C
2
then obviously both C
1
and C
2
contain the element e. The elimination property implies that
C
1
C
2
e contains a circuit. Since C
1
C
2
e is however contained in I, it is independent
(
). So I + e contains at most one circuit.

(viii)(ix): Let us denote by I the output of the greedy algorithmfor the weighting function
. (The problem is nding a heaviest independent set.) If I is a heaviest independent set, then
the matter is clear. Otherwise we take a heaviest independent set having the largest intersection
with I. Let us denote this heaviest independent set by I
. I cannot be a subset of I
, because the
greedy algorithmwould then nd an even heavier independent set. Let us further denote by e the
rst element of the set difference I I
that the greedy algorithm chooses. I
+e is a dependent
set and contains thus exactly one circuit C (remember the property of induced circuits). This
circuit of course is not included in I, so there is an element f C I. Since I
+ e contains
only one circuit, I
+ e f is an independent set. I
is maximal, so that (f) (e). On the

other hand, f and those elements of I that the greedy algorithm chose before choosing e are all
in I
, whence adding f to the elements does not create a circuit. This means that f was available
for the greedy algorithm when it chose e, and so (f) (e). We conclude that (f) = (e)
and the sets I
+ e f and I
have equal weight. This however is contrary to the choice of I
because #((I
+ e f) I) > #(I
I). (The reader may notice a similarity to the proof of

Theorem 5.2. Indeed, this gives another proof for Kruskals Algorithm No. 1.)
(ix)(iii): Let us consider independent sets I
1
and I
2
such that #(I
1
) < #(I
2
). For brevity
we denote k = #(I
1
). Consider then the weighting function
: (e) =
_
_
k + 2, if e I
1
k + 1, if e I
2
I
1
0 otherwise.
The weight of I
2
is then
eI
2
(e) (k + 1)
2
> k(k + 2) =
eI
1
(e).
It is thus larger than the weight of I
1
, so I
1
is not a heaviest independent set. On the other
hand, when nding a heaviest independent set the greedy algorithm will choose all elements of
I
1
before it ever chooses an element of I
2
I
1
. Since it is now assumed to produce a heaviest
independent set, it must choose at least one element e of I
2
I
1
and I
1
+e is thus an independent
set. This shows that the augmentation property holds true.
The most popular aspect dening a matroid is probably the augmentation property.
7.6 Operations on Matroids
In the preceding chapters, in connection with fundamental cut sets and fundamental circuits,
mutual duality was mentioned. Duality is a property that is very natural for hereditary systems
and matroids.
The dual (system) M
of a hereditary system M of the set E is a hereditary system of E

whose bases are the complements of the bases of M (against E). Often the bases of M
are
called cobases of M, circuits of M
are called cocircuits of M, and so on. It is easily checked

that M
really is a hereditary system of E: If B

1
and B
2
are distinct bases of M
then B
1
and
B
2
are distinct bases of M. Thus, if B
1
B
2
then B
2
B
1
(
). Note also that (M
= M.
Theorem 7.4. (Whitneys Theorem) The dual M
of a matroid M is a matroid, the so-called

dual matroid, and
M
(F) = #(F)
M
(E) +
M
( F ).
(Note that
M
(E) is the size of a basis of M.)
Proof. Let us show that M
has the basis exchange property, which makes it a matroid accord-

ing to Theorem 7.3. If B
1
and B
2
are distinct bases of M
and e B
1
B
2
then B
1
and B
2
are distinct bases of M and e B
2
B
1
. Since B
1
is a basis of M, B
1
+ e contains exactly
one circuit C of M (the property of induced circuits) and this circuit must have an element
f B
2
B
1
. Then however B
1
+ef does not contain a circuit of M, i.e. it is an independent
set of M, and has the same size as B
1
. All bases have the same size, so B
1
+e f is a basis of
M and its complement B
1
e + f is a basis of M
.
To compute the rank
M
(F) we take a maximal independent set H of M
included in F.
Then
M
(F) =
M
(H) = #(H).
Then H is a minimal set containing the set F and a basis of M. (This is simply the same
statement in other words. Note that H is included in some basis of M
.) But such a set is

obtained starting from F, taking a maximal independent set of M contained in Fwhich has
M
( F ) elementsand extending it to a basiswhich has
M
(E) elements. So
#( H ) #( F ) =
M
(E)
M
( F ).
Set theory tells us that
#( H ) + #(H) = #(E) = #( F ) + #(F).
Combining these we get the claimed formula for
M
(F) (check!).
Dualism gives a connection between bases of a matroid M and circuits of its dual matroid
M
(i.e. cocircuits of M):

Theorem 7.5. (i) Circuits of the dual matroid of a matroid M are the minimal sets that intersect
every basis of M.
(ii) Bases of a matroid M are the minimal sets that intersect every circuit of the dual matroid
M
.
Proof. (i) The circuits of M
are the minimal sets that are not contained in any complement of
a basis of M. Thus they must intersect every basis of M.
(ii) Bases of M
are the maximal sets that do not contain any circuit of M
. The same in
other words: Bases of M are the minimal sets that intersect every circuit of M
.
Example. Bases of the circuit matroid M(G) of a connected graph G are the spanning trees.
Bases of the dual matroid M
(G) are the complements of these, i.e. the cospanning trees. By the
theorem, circuits of the dual matroid are the cut sets of G. (Cf. Theorems 2.4 and 2.5.) Because
according to Whitneys Theorem M
(G) is a matroid, it has the greediness property, that is, the

greedy algorithm nds a heaviest/lightest basis. Working of Kruskals Algorithm No. 3 is based
on this. The algorithm nds the heaviest cospanning tree.
Analogous concepts can naturally be dened for a general, possibly disconnected, graph G.
Bases of M
(G) are then the cospanning forests of G. The dual matroid M
(G) is called the

bond matroid or the cut matroid or the cocircuit matroid of G. So, when is the bond matroid
M
(G) graphic, i.e. the circuit matroid of a graph? The so-called Whitney Planarity Theorem
tells us that this happens exactly when G is a planar graph! (See e.g. WEST.)
If M
i
is a hereditary system of the set E
i
for i = 1, . . . , k then the direct sum M = M
1

M
k
of the systems M
1
, . . . , M
k
is the hereditary system of the set E = E
1
E
k
whose
independent sets are exactly all sets I
1
I
k
where I
i
J
M
i
(i = 1, . . . , k). In particular,
if E
1
= = E
k
= E then the direct sum M is called the union of the systems M
1
, . . . , M
k
,
denoted by M = M
1
M
k
. Note that each hereditary system M
i
could also be thought of
as a hereditary system of the set E simply by adding elements of E E
i
as circuits (loops, that
is).
It is not exactly difcult to see that if M
1
, . . . , M
k
are matroids and the sets E
1
, . . . , E
k
are
pairwise disjoint then M = M
1
M
k
is a matroid, say, by demonstrating the augmentation
property (try it!). But actually a more general result holds true:
Theorem 7.6. (Matroid Union Theorem
6
) If M
1
. . . , M
k
are matroids of the set E then the
union M = M
1
M
k
is also a matroid of E and
M
:
M
(F) = min
F
F
_
#(F F
) +
k
i=1
M
i
(F
)
_
.
Proof. The proof is rather long and difcult, and is not given here (see e.g. WEST or OXLEY.)
It might be mentioned, though, that the rank formula is not valid for hereditary systems in
general.
The theorem has many fundamental corollaries, e.g.
Corollary. (Matroid Covering Theorem
7
) If M is a loopless matroid of the set E then the
smallest number of independent sets whose union equals E is
max
FE
_
#(F)
M
(F)
_
.
Proof. Note rst that since M is loopless, each element of E is in itself an independent set. The
set E thus can be covered as stated. Take nowk copies of M as the matroids M
1
, . . . , M
k
in the
union theorem. Then E is a union of k independent sets of M exactly when it is an independent
set of the union matroid M
= M
1
M
k
. The covering property we are interested in can
then be expressed in the form
M
(E) = #(E) or, by the union theorem,
#(E) = min
FE
_
#(E F) +
k
i=1
M
i
(F)
_
i.e.
min
FE
(k
M
(F) #(F)) = 0.
Since the difference to be minimized is = 0 when F is the empty set, k will be the smallest
number such that k #(F)/
M
(F) for all nonempty subsets F E.
6
Also known by the names EdmondsFulkerson Theorem and Matroid Sum Theorem.
7
Also known as Edmonds Covering Theorem.
Example. For the circuit matroid M(G) of a loopless graph G independent sets are the sub-
forests of G, and we are interested in the minimum number of subforests needed to contain all
edges of G. Let us denote this number by A(G), it is called the arboricity of G.
To analyze the maximization in the covering theorem we divide the subgraph F induced
by the edges in F into its components. Numbers of vertices and edges of these components are
denoted by n
1
, . . . , n
k
F
and m
1
, . . . , m
k
F
, respectively. We use an indexing such that
m
k
F
n
k
F
1

m
k
F
1
n
k
F
1
1

m
1
n
1
1
.
Now, in general if
x
2
y
2
x
1
y
1
then
x
2
y
2
x
1
+ x
2
y
1
+ y
2
. Thus
m
2
n
2
1

m
1
+ m
2
n
1
+ n
2
2
,
and continuing inductively, also
m
i
n
i
1

m
1
+ + m
i
n
1
+ + n
i
i
(i = 1, . . . , k
F
).
In particular then
m
k
F
n
k
F
1

m
1
+ + m
k
F
n
1
+ + n
k
F
k
F
=
#(F)
M(G)
(F)
.
Maximization can thus be restricted to edge-sets F such that F is connected and
M(G)
(F) =
n
F
1 where n
F
is the number of vertices of F. (It might be further restricted to edge-sets F
such that F also equals the subgraph induced by its vertices, since connecting two vertices by
an edge increases the numerator of the fraction to be maximized, the denominator remaining
the same.) Thus we get the celebrated Nash-Williams Formula for arboricity:
A(G) = max
FE
_
#(F)
n
F
1
_
.
It might be noted that since for a simple planar graph #(F) 3n
F
6 (Linear Bound
applied to F), A(G) is then at most 3.
The restriction of a hereditary system M of the set E into the set F E is a hereditary
system M[F whose independent sets are exactly those subsets of F that are independent sets
of M. The contraction of M into the set F is the hereditary system (M
[F)
, often denoted by
M.F. Clearly the augmentation property of M is directly transferred to M[F, so (cf. Whitneys
Theorem)
Theorem 7.7. If M is a matroid of the set E and F E then M[F and M.F are bot matroids,
too.
The minors of a matroid M are all those matroids that can be obtained from M by consecutive
restrictions and contractions.
References
1. ANDRSFAI, B.: Introductory Graph Theory. The Institute of Physics (1978)
2. ANDRSFAI, B.: Graph Theory: Flows, Matrices. The Institute of Physics (1991)
3. BANG-JENSEN, J. & GUTIN, G.: Digraphs: Theory, Algorithms and Applications. Sprin-
gerVerlag (2002)
4. BOLLOBS, B.: Modern Graph Theory. SpringerVerlag (2002)
5. CHRISTOFIDES, N.: Graph Theory. An Algorithmic Approach. Academic Press (1975)
6. DIESTEL, R.: Graph Theory. SpringerVerlag (2005)
7. DOLAN, A. & ALDOUS, J.: Networks and Algorithms. An Introductory Approach. Wiley
(1999)
8. GIBBONS, A.: Algorithmic Graph Theory. Cambridge University Press (1987)
9. GIBBONS, A. & RYTTER, W.: Efcient Parallel Algorithms. Cambridge University Press
(1990)
10. GONDRAN, M. & MINOUX, M.: Graphs and Algorithms. Wiley (1986)
11. GRIMALDI, R.P.: Discrete and Combinatorial Mathematics. AddisonWesley (2003)
12. GROSS, J. & YELLEN, J.: Graph Theory and Its Applications. CRC Press (2006)
13. GROSS, J. & YELLEN, J.: Handbook of Graph Theory. CRC Press (2003)
14. HOPCROFT, J.E. & ULLMAN, J.D.: Introduction to Automata Theory, Languages, and
Computation. AddisonWesley (1979)
15. JUNGNICKEL, D.: Graphs, Networks and Algorithms. SpringerVerlag (2004)
16. MCELIECE, R.J. & ASH, R.B. & ASH, C.: Introduction to Discrete Mathematics.
McGrawHill (1990)
17. MCHUGH, J.A.: Algorithmic Graph Theory. PrenticeHall (1990)
18. MEHLHORN, K.: Graph Algorithms and NP-Completeness. SpringerVerlag (1984)
19. NOVAK, L. & GIBBONS, A.: Hybrid Graph Theory and Network Analysis. Cambridge
University Press (1999)
20. OXLEY, J.G.: Matroid Theory. Oxford University Press (2006)
106
107
21. READ, R.C. & WILSON, R.J.: An Atlas of Graphs. Oxford University Press (2004)
22. SKIENA, S.S.: The Algorithm Design Manual. SpringerVerlag (1998)
23. SWAMY, M.N.S. & THULASIRAMAN, K.: Graphs, Networks, and Algorithms. Wiley
(1981)
24. SWAMY, M.N.S. & THULASIRAMAN, K.: Graphs: Theory and Algorithms. Wiley (1992)
25. VG, I.: Graph Theory. Application to the Calculation of Electrical Networks. Elsevier
(1985)
26. WALTHER, H.: Ten Applications of Graph Theory. Kluwer (1985)
27. WEST, D.B.: Introduction to Graph Theory. PrenticeHall (1996)
108
Index
across-quantity 43
across-source 43
across-vector 43
acyclic directed graph 32
adjacency matrix 34
adjacent edges 2
adjacent vertices 2
admittance matrix 46
all-vertec incidence matrix 34
alternating path 76
annealing algorithm 72,91
approximation algorithm 50
arboricity 105
arc 27
articulation vertex 14
aspect 92
augmentation 95,100
augmenting path 76,82
augmenting tree 77
back edge 54,56
basis 92
basis exchange property 94,100
BFS tree 59
big-O notation 50
binary matroid 96
bipartite graph 17,76,97
block 15
bond matroid 104
branch 21
Breadth-First Search 59
capacity 80
capacity constraint 80
chord 20
chromatic number 89
circuit 6,23,40,92
circuit matrix 40
circuit matroid 93,105
circuit space 49
clique 5
closed walk 6
cobasis 103
cocircuit 103
cocircuit matroid 104
coloring of a graph 89
complement of graph 10
complete bipartite graph 17
complete graph 3
component 7,28,43
computational complexity 50
condensed graph 28
connected digraph 28
connected graph 7
contracting of edge 13
contraction of matroid 105
cospanning tree 20
cross edge 56
cut 16
cut matrix 36
cut matroid 104
cut set 16,24,36
cut space 49
cut vertex 14
DavidsonHarel Algorithm 90
decision problem 50
degree of vertex 2
Demoucrons Algorithm 87
DemoucronMalgrangePertuiset Algorithm 87
dependent set 92
Depth-First Search 53
deterministic algorithm 50
DFS forest 57
DFS tree 54
difference of graphs 11
digraph 27
Dijkstras Algorithm 61
direct sum 104
directed edge 27
directed graph 27
directed spanning tree 31
directed tree 29
directed walk 27
dual hereditary system 103
dual matroid 102
edge 1
Edmonds Covering Theorem 104
EdmondsFulkerson Theorem 104
EdmondsKarp Modication 84
elimination property 95,100
empty graph 2
end vertex 2
Eulers Polyhedron Formula 86
Five-Color Theorem 89
ow 80
Floyds Algorithm 63
FordFulkerson Algorithm 83
forest 20
forward edge 56
Four-Color Theorem 89
free matroid 97
fundamental circuit 23
109
fundamental circuit matrix 41
fundamental cut set 24
fundamental cut set matrix 39
fundamental equations 44
fundamental set of circuits 23
fundamental set of cut sets 24
graph 1
graphic matroid 93
greediness property 99,100
greedy algorithm 98
Halls Theorem 79
Hamiltonian circuit 61,98
Heawoods Algorithm 90
Heawoods Theorem 89
hereditary family 92
hereditary set 92
HopcroftTarjan Algorithm 87
Hungarian Algorithm 77
Hungarian tree 77
impedance matrix 46
in-degree 27
incidence matrix 35
independent set 92
induced subgraph 5
intersection of graphs 11
intractable problem 51
isolated vertex 2
isomorphic graphs 18
Jarniks Algorithm 70
KarpHeld Heuristics 73
Kirchhoffs Across-Quantity Law 43
Kirchhoffs Flow Law 80
Kirchhoffs Through-Quantity Law 43
Kruskals Algorithm 67,98,104
Kuratowskis Theorem 87
labeled graph 18
labeling 18
Las Vegas algorithm 51
leaf 29
lightest Hamiltonian circuit 71
lightest path 61,63
lightest spanning tree 66
Linear Bound 86,105
linear matroid 96
link 21
loop 2,92
Marimonts Algorithm 33
Marriage Theorem 79
matching 76,97
matrix matroid 96
matroid 100
Matroid Covering Theorem 104
Matroid Greediness Theorem 99
Matroid Sum Theorem 104
Matroid Union Theorem 104
Max-Flow Min-Cut Theorem 83
maximal matching 76
maximum degree 3
maximum matching 76,84
minimum degree 3
Minimum Degree Bound 87
minor 105
Monte Carlo algorithm 51
multiplicity 1,12
multiset 1
NasWilliams Formula 105
NP 51
NP-complete 51,71
NP-hard 51,91
nondeterministic algorithm 50
null graph 2
nullity of graph 8
open walk 6 out-degree 27
P 51
parallel edges 2
parallel elements 92
partition matroid 97
path 6
pendant edge 2
pendant vertex 2
perfect matching 79
planae embedding 85
planar graph 85,104,105
polynmial time 51
polynomial space 51
potential vector 43
Prims Algorithm 70
probabilistic algorithm 51
proper difference 12
property of induced circuits 96,100
quasi-strongly connected digraph 29
rank function 93
rank of graph 8
rank of matroid 93
reachability matrix 52
reference vertex 35
region 85
removal of edge 13
removal of vertex 12
representation 96
restriction of matroid 105
ring sum of graphs 11,23
root 29
separable graph 14
110
short-circuiting of vertices 13
shortest path 61
simple graph 2
spanning tree 20
stationary linear network 43
stochastic algorithm 51
strong absorptivity 95,100
strongly connected 28
strongly connected component 28
subforest 20
subgraph 3
submodularity 100
subtree 20
symmetric difference 11
Tellegens Theorem 48
through-quantity 43
through-source 43
through-vector 43
topological sorting 32
tractable problem 51
trail 6
transport network 80
transversal matroid 97
Travelling Salesmans Problem 71
tree 20,29
tree edge 54,56,59
trivial graph 2
underlying graph 27
uniform matroid 97
uniformity 94,100
union of graphs 11
union of matroids 104
vectorial matroid 96
vertex 1
walk 6
Warshalls Algorithm 52
weak absorptivity 94,100
weights 18
Whitneys Planarity Theorem 104
Whitneys Theorem 103

GT English

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

GT English

Uploaded by

Copyright:

Available Formats

GRAPH THEORY

is a vertex of G such that v

is connected to some vertex of V

. We have two cases:

= k. We use the Induction Hypothesis on G

A graph is circuitless if it does not have any circuit in it.

), where k . We choose the smallest index t such that

means contradiction. If we get a contradiction by proceeding from the assump-

. Then, there would be two parallel edges between two

(Verify the case t = s = r.)

), or k 2 and the two vertices v

) are simple and V

contains those edges from G that are not

). The same reasoning can be used for all the uw paths in G.

is the cospanning tree.

). On the other hand, if all but one vertex have

). (This also follows from

Next, we consider some properties of circuits and cut sets:

has a circuit and T is not a tree.

. Thus, (ii) is true.

in C. Since (ii) is true, C

. The other edges

in C. By Theorem 2.6, we know that the common edges of C and I

). So c is only in the fundamental cut sets I

whose corresponding fundamental circuit

). Hence, the branch b is only

is the initial vertex and v

of G where n 2. The following process shows that there must be a root:

) We still have to show that d

(r) = 0. Let us consider the

(r) 1 or the in-degree of some vertex on that

is the minor corresponding to a

4.3 Circuit Matrix

. The circuit matrix of G is an m matrix B = (b

come last and sort the fundamental

0 if vertex x has not been visited

) (the other case is obviously symmetrical).

. Obviously, DFS goes through the vertices of T

is completely scanned only after all the vertices of T

The corresponding DFS forest is

1 if the label is permanent

, a directed uv path does not exist and we stop.

in step #3 to be some predecessor of s on the directed uws path). Then, we get a

the predecessor of w on the lightest directed uw path if there is one and w ,= u

is on the lightest directed path but it is not v

whose vertices other than v

path of this kind has common vertices

). (We have to remember that weights of

is the lightest di-

path which contains the vertices v

. Now, we use the Induction Hypothesis:

is not on the path. Now, we construct the

path and the lightest v

path with the v

is the vertex following v

. The spanning tree T is locally minimal if (T) (T

. Other than c, the branches in C are

has (exactly one) branch e

which is not in T, i.e. e

. Not all edges of C are in T

. Then, e is a branch of T (actually the only branch of T that is not in T

to T. The result has to be T

. Because of circuit minimality,